FLUX.2 [klein] 4B

Exact model and requirements

Artifact:

Black Forest Labs / FLUX.2 Klein4B distilled complete Diffusers bundle · checkpoint black-forest-labs/FLUX.2-klein-4B

Publisher revision:

e7b7dc27f91deacad38e78976d1f2b499d76a294

Artifact digest:

A multi-file pipeline has no single publisher aggregate byte digest. Each of19 required files is independently bound in the reader manifest by publisher LFS SHA256 or exact Git blob SHA1.

Download size:

14.87 GiB

Memory requirement (unknown):

Unknown — No measured full-process peak for full FP32 transformer/text encoder/VAE plus512image activations.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host, one prompt, one image, one serial request. All complete native modules remain on CPU float32; no GPU or disk offload.

KV cache:

No autoregressive text generation or retained text K/V cache. Native Qwen3 encoder runs use_cache=False; diffusion latents are distinct.

Workload:

512 token context; 512 input tokens; Not applicable; concurrency 1; batch 1; cache cold; repetitions Unknown.

Tokenizer:

Bundle Qwen2TokenizerFast and native user chat template, thinking disabled;512 padded prompt positions · e7b7dc27f91deacad38e78976d1f2b499d76a294

Sampling / reasoning:

Fixed original prompt: a red cube beside a blue sphere on white. Complete formatted prompt is counted before native truncation and rejected above 512. Native pipeline pads to 512 positions, uses Qwen3 hidden layers 9/18/27 with no text generation/cache. One512 × 512 image, four distilled steps, guidance_scale 1.0, CPU generator seed 0. Seed is a reproducibility parameter, not a cross-hardware image equality guarantee. 512 is the chosen native prompt-encoder maximum including template; the unpadded count is reported at reader execution. It is not the Qwen3 theoretical 40960 capacity or total joint image/text attention. VAE downsampling8 and2 × 2 latent packing give 1024 image latent positions at 512square, separate from text. Four steps are the publisher distilled recipe, not a text sampler.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

The actual bundle LICENSE.md and publisher model card grant FLUX.2 Klein4B under Apache License2.0, including this selected complete bundle. Preserve applicable license/copyright/NOTICE and change/patent conditions. Do not transfer noncommercial 9B or FLUX.2 dev terms from names in VAE metadata to this 4B bundle. Publisher encourages safety filtering; its card says recommendations do not modify the grant. No safety classifier is claimed here; fixed benign geometry prompt, manual inspection and applicable laws still matter. This is no legal clearance or trademark endorsement.

• text_encoder/model-00001-of-00002.safetensors · 4.63 GiB · revision e7b7dc27f91deacad38e78976d1f2b499d76a294
• text_encoder/model-00002-of-00002.safetensors · 2.87 GiB · revision e7b7dc27f91deacad38e78976d1f2b499d76a294
• transformer/diffusion_pytorch_model.safetensors · 7.22 GiB · revision e7b7dc27f91deacad38e78976d1f2b499d76a294
• vae/diffusion_pytorch_model.safetensors · 0.16 GiB · revision e7b7dc27f91deacad38e78976d1f2b499d76a294

• 15,964,212,614 bytes is the sum of four official safetensors file sizes, not full installed footprint or runtime RAM. The standalone7.75 GB transformer file alone is not the complete pipeline.
• 64 GiB RAM, 80 GiB disk and eight threads are planning choices. CPU float32 increases weight storage relative to source tensor dtype; image attention, decoder, temporary load buffers, packages and OS add unknown peak memory.

Install the runtime

1. 1. Install the compatible native CPU stack

   Use a dedicated virtualenv and official matched CPU wheel hashes. Diffusers 0.37.1 is chosen because its HFhub range intersects Transformers 4.57.6; latest Diffusers 0.41.0 requires a different hub generation and is not substituted. No all-extra, CUDA, offload, optional flash kernel or remote code is enabled.

   CommandCopy command

   set -e
   sudo apt-get update
   sudo apt-get install -y python3.12 python3.12-venv curl ca-certificates openssl
   mkdir -p "$HOME/flux2-klein4b-image/model"
   cd "$HOME/flux2-klein4b-image"
   python3.12 -m venv .venv
   .venv/bin/python -m pip install 'https://download.pytorch.org/whl/cpu/torch-2.6.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=59e78aa0c690f70734e42670036d6b541930b8eabbaa18d94e090abf14cc4d91' 'https://download-r2.pytorch.org/whl/cpu/torchvision-0.21.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=d6874431e678ba107b60a83f255c33f3755f06bad587b1b919aa514ec325dcd8'
   .venv/bin/python -m pip install 'diffusers==0.37.1' 'transformers==4.57.6' 'tokenizers==0.22.2' 'huggingface-hub==0.36.2' 'accelerate==1.12.0' 'safetensors==0.7.0' 'numpy==2.2.3' 'Pillow==11.1.0'
   .venv/bin/python -m pip check
   .venv/bin/python -m pip freeze > resolved-environment.txt

   Expected result: Pinned direct versions resolve and pip check passes; exact transitive/system resolution is saved. No native execution/compatibility or fit claim.

Download the model

1. 1. Write the complete immutable component manifest

   This reader-authored manifest selects 19 exact bundle files, including all four weight files. LFS blobs use publisher SHA256; regular source files use their exact Git blob SHA1. SHA1 is Git identity, not a modern adversarial integrity guarantee; immutable official HTTPS revision remains the provenance boundary.

   CommandCopy command

   set -e
   cd "$HOME/flux2-klein4b-image"
   cat > component-manifest.json <<'MANIFEST'
   {
     "repository": "black-forest-labs/FLUX.2-klein-4B",
     "revision": "e7b7dc27f91deacad38e78976d1f2b499d76a294",
     "files": [
       {
         "file": "LICENSE.md",
         "bytes": 9584,
         "algorithm": "git-blob-sha1",
         "digest": "0afa7632c1a195ea6627a3c53e6ebf18d1be3aa5"
       },
       {
         "file": "model_index.json",
         "bytes": 446,
         "algorithm": "git-blob-sha1",
         "digest": "554a39bf1019138dd0b28809b745df8ef33657c8"
       },
       {
         "file": "scheduler/scheduler_config.json",
         "bytes": 486,
         "algorithm": "git-blob-sha1",
         "digest": "7e53605e2643440b310fa554f087641893883c5f"
       },
       {
         "file": "text_encoder/config.json",
         "bytes": 1536,
         "algorithm": "git-blob-sha1",
         "digest": "86f951d170cfd1fdc1c8a57a041e05c27b7b789b"
       },
       {
         "file": "text_encoder/generation_config.json",
         "bytes": 214,
         "algorithm": "git-blob-sha1",
         "digest": "f5af93d0c7764fa38e3b1cd3aaba1b6307246d3b"
       },
       {
         "file": "text_encoder/model-00001-of-00002.safetensors",
         "bytes": 4967215360,
         "algorithm": "sha256",
         "digest": "8c0506e7f4936fa7e26183a4fd8da4e2bdbc5990ba64ae441f965d51228f36ea"
       },
       {
         "file": "text_encoder/model-00002-of-00002.safetensors",
         "bytes": 3077766632,
         "algorithm": "sha256",
         "digest": "82f2bd839378541b0557bfabaf37c7d3d637071fdcb73302dedd7cf61162ce07"
       },
       {
         "file": "text_encoder/model.safetensors.index.json",
         "bytes": 32855,
         "algorithm": "git-blob-sha1",
         "digest": "b65d8063815c8679d85bb919337b36fae8de75cf"
       },
       {
         "file": "tokenizer/added_tokens.json",
         "bytes": 707,
         "algorithm": "git-blob-sha1",
         "digest": "b54f9135e44c1e81047e8d05cb027af8bc039eed"
       },
       {
         "file": "tokenizer/chat_template.jinja",
         "bytes": 4168,
         "algorithm": "git-blob-sha1",
         "digest": "01be9b307daa2d425f7c168c9fb145a286e0afb4"
       },
       {
         "file": "tokenizer/merges.txt",
         "bytes": 1671853,
         "algorithm": "git-blob-sha1",
         "digest": "31349551d90c7606f325fe0f11bbb8bd5fa0d7c7"
       },
       {
         "file": "tokenizer/special_tokens_map.json",
         "bytes": 613,
         "algorithm": "git-blob-sha1",
         "digest": "ac23c0aaa2434523c494330aeb79c58395378103"
       },
       {
         "file": "tokenizer/tokenizer.json",
         "bytes": 11422654,
         "algorithm": "sha256",
         "digest": "aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4"
       },
       {
         "file": "tokenizer/tokenizer_config.json",
         "bytes": 5404,
         "algorithm": "git-blob-sha1",
         "digest": "ddaf69808214a44fdd26d3785b66c1367c78277a"
       },
       {
         "file": "tokenizer/vocab.json",
         "bytes": 2776833,
         "algorithm": "git-blob-sha1",
         "digest": "4783fe10ac3adce15ac8f358ef5462739852c569"
       },
       {
         "file": "transformer/config.json",
         "bytes": 541,
         "algorithm": "git-blob-sha1",
         "digest": "f21592ae2bc9f43052e78785bde7414f219ae369"
       },
       {
         "file": "transformer/diffusion_pytorch_model.safetensors",
         "bytes": 7751109744,
         "algorithm": "sha256",
         "digest": "9f29f9edcfdae452a653ffb51a534ca4decd389952c225724ff3b94042612a6e"
       },
       {
         "file": "vae/config.json",
         "bytes": 821,
         "algorithm": "git-blob-sha1",
         "digest": "c3f38eb4c188a96a462519159356bd0fcc28cd14"
       },
       {
         "file": "vae/diffusion_pytorch_model.safetensors",
         "bytes": 168120878,
         "algorithm": "sha256",
         "digest": "ca70d2202afe6415bdbcb8793ba8cd99fd159cfe6192381504d6c4d3036e0f04"
       }
     ]
   }
   MANIFEST
   sha256sum --check <<'SHA256'
   f03107bcd53ab215deeee44072681f15658a94a755ae242bb42741dfca6405e3  component-manifest.json
   SHA256

   Expected result: Exact complete manifest verifies before any reader weight retrieval.

2. 2. Retrieve and verify all native components

   Reader-only downloads from one exact ungated publisher revision. Do not fetch the standalone transformer-only file or float revisions. Stop on any size/hash mismatch; no model loading occurs in this retrieval script.

   CommandCopy command

   set -e
   cd "$HOME/flux2-klein4b-image"
   .venv/bin/python - <<'PYDOWNLOAD'
   import hashlib,json,pathlib,shutil
   from huggingface_hub import hf_hub_download
   m=json.load(open('component-manifest.json'))
   for row in m['files']:
    src=hf_hub_download(m['repository'],row['file'],revision=m['revision'])
    dst=pathlib.Path('model')/row['file'];dst.parent.mkdir(parents=True,exist_ok=True);shutil.copyfile(src,dst)
    assert dst.stat().st_size==row['bytes']
    h=hashlib.sha256() if row['algorithm']=='sha256' else hashlib.sha1()
    if row['algorithm']=='git-blob-sha1':h.update(('blob '+str(row['bytes'])+'\0').encode())
    with dst.open('rb') as f:
     for chunk in iter(lambda:f.read(1024*1024),b''):h.update(chunk)
    assert h.hexdigest()==row['digest'],row['file']
   print('All19 components verified; model loading has not run.')
   PYDOWNLOAD

   Expected result: All 19 exact files verify. Download cache duplicates disk usage;80 GiB is an allocation, not a measured peak. Offline loading follows only after complete verification.

Configure

1. 1. Create a stable private service key

   Generate only when absent; preserve existing access. Do not print or share the local key.

   CommandCopy command

   set -e
   cd "$HOME/flux2-klein4b-image"
   umask 077
   if [ ! -e .api-key ]; then openssl rand -hex 32 > .api-key; fi
   chmod 600 .api-key
   test -s .api-key || { echo "Restore empty key before continuing"; exit 1; }

   Expected result: Private64-character lowercase hexadecimal key preserved or created.

2. 2. Create the native image harness

   Reader-authored wrapper, not publisher server. It rechecks every component before explicit native local_files_only safetensors component loading. Reject any missing, unexpected, mismatched or error diagnostic before assembling the native pipeline. All modules use CPU float32; fixed prompt is chat-formatted with thinking disabled and checked before native 512 truncation. No full model binary was fetched or executed by the author.

   CommandCopy command

   set -e
   cd "$HOME/flux2-klein4b-image"
   cat > flux2_image.py <<'PYSERVER'
   """Reader-authored educational fixed-input FLUX.2 Klein4B harness; not publisher server."""
   import hashlib
   import hmac
   import json
   import math
   import pathlib
   from http.server import BaseHTTPRequestHandler, HTTPServer
   
   ROOT = pathlib.Path(__file__).resolve().parent
   PORT = 8995
   
   
   PROMPT = 'A simple red cube beside a blue sphere on a plain white background, clean studio illustration.'
   
   
   def validate_image(result):
       import base64
       import struct
       png, count = result
       if type(count) is not int or not 0 < count <= 512:
           raise ValueError('prompt_budget')
       if not isinstance(png, bytes) or not 57 <= len(png) <= 8*1024*1024:
           raise ValueError('png_bound')
       if png[:8] != b'\x89PNG\r\n\x1a\n' or png[12:16] != b'IHDR':
           raise ValueError('png_header')
       if struct.unpack('>II', png[16:24]) != (512, 512) or png[-8:-4] != b'IEND':
           raise ValueError('png_shape_or_end')
       return {'fixture': 'author-t2i-v1', 'prompt': PROMPT,
               'prompt_tokens_before_padding': count, 'padded_prompt_positions': 512,
               'image_shape': [512, 512, 3], 'configured_image_latent_positions': 1024,
               'denoising_steps': 4, 'seed': 0, 'guidance_scale': 1.0,
               'mime_type': 'image/png', 'png_base64': base64.b64encode(png).decode('ascii'),
               'generated_text_tokens': None}
   
   
   def make_encoder():
       # Reader-run only. Fake tests do not import Torch, Diffusers or Pillow.
       import os
       os.environ['HF_HUB_OFFLINE'] = '1'
       os.environ['TRANSFORMERS_OFFLINE'] = '1'
       import torch
       from diffusers import (Flux2KleinPipeline, Flux2Transformer2DModel,
                              AutoencoderKLFlux2, FlowMatchEulerDiscreteScheduler)
       from transformers import Qwen3ForCausalLM, Qwen2TokenizerFast
       manifest = json.loads((ROOT / 'component-manifest.json').read_text())
       model_root = ROOT / 'model'
       for row in manifest['files']:
           file = model_root / row['file']
           if file.stat().st_size != row['bytes']:
               raise RuntimeError('component_size')
           hasher = hashlib.sha256() if row['algorithm'] == 'sha256' else hashlib.sha1()
           if row['algorithm'] == 'git-blob-sha1':
               hasher.update(('blob '+str(row['bytes'])+'\0').encode())
           with file.open('rb') as stream:
               for chunk in iter(lambda: stream.read(1024*1024), b''):
                   hasher.update(chunk)
           if hasher.hexdigest() != row['digest']:
               raise RuntimeError('component_hash')
       torch.set_num_threads(8)
       def load_checked(cls, subfolder):
           component, info = cls.from_pretrained(
               str(model_root / subfolder), torch_dtype=torch.float32,
               local_files_only=True, use_safetensors=True, output_loading_info=True)
           for key in ['missing_keys', 'unexpected_keys', 'mismatched_keys', 'error_msgs']:
               if info.get(key):
                   raise RuntimeError('component_loading_diagnostics_'+subfolder+'_'+key)
           return component.float().cpu().eval()
       transformer = load_checked(Flux2Transformer2DModel, 'transformer')
       text_encoder = load_checked(Qwen3ForCausalLM, 'text_encoder')
       vae = load_checked(AutoencoderKLFlux2, 'vae')
       tokenizer = Qwen2TokenizerFast.from_pretrained(str(model_root / 'tokenizer'), local_files_only=True)
       scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
           str(model_root / 'scheduler'), local_files_only=True)
       pipe = Flux2KleinPipeline(transformer=transformer, text_encoder=text_encoder,
                                 vae=vae, tokenizer=tokenizer, scheduler=scheduler,
                                 is_distilled=True)
       pipe = pipe.to('cpu')
       pipe.set_progress_bar_config(disable=True)
       if not json.loads((model_root / 'model_index.json').read_text())['is_distilled'] or not pipe.config.is_distilled:
           raise RuntimeError('wrong_variant')
       formatted = pipe.tokenizer.apply_chat_template(
           [{'role': 'user', 'content': PROMPT}], tokenize=False,
           add_generation_prompt=True, enable_thinking=False)
       token_count = len(pipe.tokenizer(formatted, truncation=False)['input_ids'])
       if not 0 < token_count <= 512:
           raise RuntimeError('prompt_would_truncate')
   
       def encode():
           import io
           import numpy as np
           from PIL import Image
           def check_latents(pipeline, step, timestep, values):
               if not torch.isfinite(values['latents']).all():
                   raise RuntimeError('nonfinite_latents')
               return values
           with torch.inference_mode():
               image = pipe(prompt=PROMPT, height=512, width=512, num_inference_steps=4,
                            guidance_scale=1.0, max_sequence_length=512,
                            num_images_per_prompt=1, output_type='np',
                            callback_on_step_end=check_latents,
                            callback_on_step_end_tensor_inputs=['latents'],
                            generator=torch.Generator(device='cpu').manual_seed(0)).images[0]
           if image.shape != (512, 512, 3) or not np.isfinite(image).all():
               raise RuntimeError('image_shape_or_finiteness')
           if image.min() < 0 or image.max() > 1:
               raise RuntimeError('image_range')
           image = Image.fromarray((image*255).round().astype('uint8'), mode='RGB')
           buf = io.BytesIO()
           image.save(buf, format='PNG')
           return buf.getvalue(), token_count
       return encode
   
   
   PYSERVER

   Expected result: First complete native loader and PNG-validation fragment written.

3. 3. Complete and verify private transport

   Append fixed-input private serial HTTP transport. No model code runs in syntax/hash checks.

   CommandCopy command

   set -e
   cd "$HOME/flux2-klein4b-image"
   cat >> flux2_image.py <<'PYSERVER'
   def create_server(encoder, key, port=PORT):
       if not isinstance(key, str) or len(key) != 64 or any(c not in '0123456789abcdef' for c in key):
           raise ValueError('key_format')
   
       class Handler(BaseHTTPRequestHandler):
           def setup(self):
               super().setup()
               self.connection.settimeout(5)
   
           def log_message(self, *_):
               pass  # Do not log Bearer keys or request bodies.
   
           def reply(self, status, obj):
               data = json.dumps(obj, allow_nan=False, separators=(',', ':')).encode()
               self.send_response(status)
               self.send_header('Content-Type', 'application/json')
               self.send_header('Content-Length', str(len(data)))
               self.send_header('Connection', 'close')
               self.end_headers()
               self.wfile.write(data)
               self.close_connection = True
   
           def authorized(self):
               values = self.headers.get_all('Authorization', [])
               return len(values) == 1 and hmac.compare_digest(values[0].encode(), ('Bearer ' + key).encode())
   
           def do_GET(self):
               if not self.authorized():
                   return self.reply(401, {'error': 'unauthorized'})
               if self.path != '/health':
                   return self.reply(404, {'error': 'not_found'})
               return self.reply(200, {'ready': True, 'padded_prompt_positions': 512, 'configured_image_latent_positions': 1024,
                                       'generated_text_tokens': None})
   
           def do_POST(self):
               if not self.authorized():
                   return self.reply(401, {'error': 'unauthorized'})
               if self.path != '/v1/fixed-image':
                   return self.reply(404, {'error': 'not_found'})
               if self.headers.get('Transfer-Encoding') is not None:
                   return self.reply(400, {'error': 'transfer_encoding_unsupported'})
               lengths = self.headers.get_all('Content-Length', [])
               if len(lengths) != 1:
                   return self.reply(411, {'error': 'content_length_required'})
               raw = lengths[0]
               if not raw or len(raw) > 4 or any(c not in '0123456789' for c in raw):
                   return self.reply(413, {'error': 'invalid_or_excessive_length'})
               n = int(raw)
               if n < 1 or n > 256:
                   return self.reply(413, {'error': 'body_limit'})
               if self.headers.get_content_type() != 'application/json':
                   return self.reply(415, {'error': 'json_required'})
               try:
                   body = self.rfile.read(n)
               except (TimeoutError, OSError):
                   return self.reply(408, {'error': 'body_timeout'})
               if len(body) != n:
                   return self.reply(400, {'error': 'incomplete_body'})
               try:
                   def unique(pairs):
                       obj = {}
                       for k, v in pairs:
                           if k in obj:
                               raise ValueError('duplicate_key')
                           obj[k] = v
                       return obj
                   request = json.loads(body.decode('utf-8'), object_pairs_hook=unique,
                                        parse_constant=lambda _: (_ for _ in ()).throw(ValueError('constant')))
               except (ValueError, UnicodeError, RecursionError):
                   return self.reply(400, {'error': 'invalid_json'})
               if type(request) is not dict or request != {'fixture': 'author-t2i-v1'}:
                   return self.reply(400, {'error': 'fixed_fixture_required'})
               try:
                   result = validate_image(encoder())
               except Exception:
                   return self.reply(500, {'error': 'pipeline_or_png_failure'})
               return self.reply(200, result)
   
       return HTTPServer(('127.0.0.1', port), Handler)
   
   
   if __name__ == '__main__':
       secret = (ROOT / '.api-key').read_text().strip()
       encoder = make_encoder()
       with create_server(encoder, secret) as server:
           server.serve_forever()
   PYSERVER
   .venv/bin/python -m py_compile flux2_image.py
   sha256sum --check <<'SHA256'
   d4e624292a3cec1f07856e464c0020b3d2db6591555e04ffe673ac3f2dffaefb  flux2_image.py
   SHA256

   Expected result: Exact wrapper bytes and Python syntax verify without model loading.

Start the model

1. 1. Start the complete CPU image experiment

   Foreground serial endpoint. Full pipeline load and first image may be slow or exceed RAM; no fit/speed is promised. Explicit native loaders select all model_index components, reject loading diagnostics and assemble the native pipeline with offline safetensors and CPU float32; no VAE/text-encoder substitution. Each denoising-step latent tensor and final NumPy image must be finite before PNG conversion.

   CommandCopy command

   set -e
   cd "$HOME/flux2-klein4b-image"
   HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 .venv/bin/python flux2_image.py

   Expected result: After actual reader-side component verification and native load, loopback 127.0.0.1:8995 serves. Stop on component/import/load/template errors.

Verify the setup

1. 1. Validate the PNG and inspect the generated image

   Run in a second server terminal. Check private health, missing-key 401, fixed response metadata and a valid decoded 512 RGB PNG. Inspect image.png yourself for the red cube/blue sphere on white; valid dimensions do not establish semantic fidelity, safety or quality. No generated example or success claim is supplied.

   CommandCopy command

   set -e
   cd "$HOME/flux2-klein4b-image"
   FLUX_KEY="$(cat .api-key)"
   curl --fail http://127.0.0.1:8995/health -H "Authorization: Bearer $FLUX_KEY"
   curl --fail http://127.0.0.1:8995/v1/fixed-image -H "Authorization: Bearer $FLUX_KEY" -H 'Content-Type: application/json' -d '{"fixture":"author-t2i-v1"}' -o image.json
   curl -s -o /dev/null -w '%{http_code}\n' http://127.0.0.1:8995/health
   unset FLUX_KEY
   .venv/bin/python - <<'PYVERIFY'
   import base64,io,json
   from PIL import Image
   d=json.load(open('image.json'));assert d['generated_text_tokens'] is None
   assert d['image_shape']==[512,512,3] and d['padded_prompt_positions']==512
   assert 0<d['prompt_tokens_before_padding']<=512 and d['configured_image_latent_positions']==1024
   assert d['denoising_steps']==4 and d['seed']==0
   b=base64.b64decode(d['png_base64'],validate=True);assert len(b)<=8*1024*1024
   with Image.open(io.BytesIO(b)) as image:
    assert image.format=='PNG' and image.size==(512,512) and image.mode=='RGB';image.verify()
   open('image.png','wb').write(b)
   print('Inspect image.png; PNG validity is not semantic quality proof.')
   PYVERIFY

   Expected result: Authenticated health/generation return structured JSON; missing auth 401. Exact PNG metadata/Pillow verify passes only after reader execution. Inspect and record actual visual defects or loading failures.

Performance evidence

No compatible measurement available. Documented support and estimated memory do not establish response speed.

Limits and troubleshooting

• Complete official 4B distilled bundle is selected, not transformer-only 7.75 GB weights or9B/base/dev. Main tensor file metadata plus all required configs/processors are pinned; no model binary was downloaded.
• 512 configured prompt positions include the native user template with thinking disabled. Actual pre-padding prompt count is checked at reader execution; too-long prompts stop before native silent truncation. This is not 40960 theoretical Qwen3 context or 1024 image latent positions.
• Native four-step distilled image synthesis returns PNG pixels, not autoregressive output tokens. Guidance_scale 1.0 avoids CFG; no chat completion/reasoning output or text-generation benchmark is claimed.
• Full CPU float32 through native .to(cpu) is source-qualified, not installed/executed compatibility proof. 64 GiB RAM, 80 GiB disk/eight threads are planning allocations; fit, peak memory, latency and image quality remain unknown.
• Original fixed geometric prompt is not a third-party image fixture. Manual inspection remains necessary; fixed seed does not guarantee equal bytes across hardware/library versions. No downstream safety classifier or commercial suitability is claimed.
• Reader-authored serial HTTPServer has private key/loopback, bounded strict fixed JSON and PNG-size/shape guards, but no inference deadline, multiuser robustness or public production-service guarantee.
• All components are verified before offline local loading; no remote-code trust or unpinned auxiliary download is selected. Publisher Git blob SHA1 is identified accurately; weight SHA256s are metadata until reader actual download.
• Fake PNG protocol tests prove only request/auth/output boundary handling, not actual native tensor finiteness, denoising correctness, prompt adherence, image safety or runtime fit. Native callbacks and pre-PNG guards require finite tensors/arrays at reader execution; they have not run on model weights here.
• Image editing, multiple references, other resolutions, base variants, alternate supported GPU/quantized/offload backends and other useful modes remain open; this pair does not finalize family coverage.

Component/hash/version or offline load error

Stop and restore exact 19-file manifest and publisher revision. Do not float to main, omit text encoder/VAE or substitute standalone transformer-only weights. Keep safetensors/local_files_only; stop on unresolved dependency errors.

Prompt budget or wrong variant error

Use exact fixed prompt/tokenizer/template and distilled model_index. The pre-padding formatted prompt must fit 512 positions. Do not hide truncation, switch model variants or treat padding count as actual semantic prompt length.

401 or rejected body

Use preserved matching private 64-character key and exact fixture selector. Arbitrary prompts/images/URLs/files or extra fields are rejected.

Slow, out-of-memory, black or inaccurate image

Record actual load/generation failure and inspect PNG. 64 GiB is not measured fit and valid dimensions do not establish semantic quality. Do not claim a smaller/partial pipeline is equivalent; stop the process if resource use is unacceptable.

Stop the model

1. 1. Stop the private image service

   Press Ctrl+C in its foreground terminal and unset client key. Confirm local loopback endpoint no longer serves.

   CommandCopy command

   unset FLUX_KEY

   Expected result: Pipeline process and endpoint stop; generated files remain until explicitly removed.

Uninstall

1. 1. Remove dedicated experiment and selected model cache

   Preserve desired outputs/audit/license records first. Confirm the dedicated directory before deletion. The model-specific download cache is separate; remove only this exact cache if no other experiment needs it. Shared Ubuntu packages remain.

   CommandCopy command

   rm -r -- "$HOME/flux2-klein4b-image"
   rm -r -- "$HOME/.cache/huggingface/hub/models--black-forest-labs--FLUX.2-klein-4B"

   Expected result: Dedicated bundle/venv/wrapper/key/outputs and this selected model cache are removed if present.

Exact model and requirements

Artifact:

Black Forest Labs / FLUX.2 Klein4B distilled complete Diffusers bundle · checkpoint black-forest-labs/FLUX.2-klein-4B

Publisher revision:

e7b7dc27f91deacad38e78976d1f2b499d76a294

Artifact digest:

A multi-file pipeline has no single publisher aggregate byte digest. Each of19 required files is independently bound in the reader manifest by publisher LFS SHA256 or exact Git blob SHA1.

Download size:

14.87 GiB

Memory requirement (unknown):

Unknown — No measured full-process peak for full FP32 transformer/text encoder/VAE plus512image activations.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host, one prompt, one image, one serial request. All complete native modules remain on CPU float32; no GPU or disk offload.

KV cache:

No autoregressive text generation or retained text K/V cache. Native Qwen3 encoder runs use_cache=False; diffusion latents are distinct.

Workload:

512 token context; 512 input tokens; Not applicable; concurrency 1; batch 1; cache cold; repetitions Unknown.

Tokenizer:

Bundle Qwen2TokenizerFast and native user chat template, thinking disabled;512 padded prompt positions · e7b7dc27f91deacad38e78976d1f2b499d76a294

Sampling / reasoning:

Fixed original prompt: a red cube beside a blue sphere on white. Complete formatted prompt is counted before native truncation and rejected above 512. Native pipeline pads to 512 positions, uses Qwen3 hidden layers 9/18/27 with no text generation/cache. One512 × 512 image, four distilled steps, guidance_scale 1.0, CPU generator seed 0. Seed is a reproducibility parameter, not a cross-hardware image equality guarantee. 512 is the chosen native prompt-encoder maximum including template; the unpadded count is reported at reader execution. It is not the Qwen3 theoretical 40960 capacity or total joint image/text attention. VAE downsampling8 and2 × 2 latent packing give 1024 image latent positions at 512square, separate from text. Four steps are the publisher distilled recipe, not a text sampler.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

The actual bundle LICENSE.md and publisher model card grant FLUX.2 Klein4B under Apache License2.0, including this selected complete bundle. Preserve applicable license/copyright/NOTICE and change/patent conditions. Do not transfer noncommercial 9B or FLUX.2 dev terms from names in VAE metadata to this 4B bundle. Publisher encourages safety filtering; its card says recommendations do not modify the grant. No safety classifier is claimed here; fixed benign geometry prompt, manual inspection and applicable laws still matter. This is no legal clearance or trademark endorsement.

• text_encoder/model-00001-of-00002.safetensors · 4.63 GiB · revision e7b7dc27f91deacad38e78976d1f2b499d76a294
• text_encoder/model-00002-of-00002.safetensors · 2.87 GiB · revision e7b7dc27f91deacad38e78976d1f2b499d76a294
• transformer/diffusion_pytorch_model.safetensors · 7.22 GiB · revision e7b7dc27f91deacad38e78976d1f2b499d76a294
• vae/diffusion_pytorch_model.safetensors · 0.16 GiB · revision e7b7dc27f91deacad38e78976d1f2b499d76a294

• 15,964,212,614 bytes is the sum of four official safetensors file sizes, not full installed footprint or runtime RAM. The standalone7.75 GB transformer file alone is not the complete pipeline.
• 64 GiB RAM, 80 GiB disk and eight threads are planning choices. CPU float32 increases weight storage relative to source tensor dtype; image attention, decoder, temporary load buffers, packages and OS add unknown peak memory.

Install the runtime

1. 1. Install the compatible native CPU stack

   Use a dedicated virtualenv and official matched CPU wheel hashes. Diffusers 0.37.1 is chosen because its HFhub range intersects Transformers 4.57.6; latest Diffusers 0.41.0 requires a different hub generation and is not substituted. No all-extra, CUDA, offload, optional flash kernel or remote code is enabled.

   CommandCopy command

   set -e
   sudo apt-get update
   sudo apt-get install -y python3.12 python3.12-venv curl ca-certificates openssl
   mkdir -p "$HOME/flux2-klein4b-image/model"
   cd "$HOME/flux2-klein4b-image"
   python3.12 -m venv .venv
   .venv/bin/python -m pip install 'https://download.pytorch.org/whl/cpu/torch-2.6.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=59e78aa0c690f70734e42670036d6b541930b8eabbaa18d94e090abf14cc4d91' 'https://download-r2.pytorch.org/whl/cpu/torchvision-0.21.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=d6874431e678ba107b60a83f255c33f3755f06bad587b1b919aa514ec325dcd8'
   .venv/bin/python -m pip install 'diffusers==0.37.1' 'transformers==4.57.6' 'tokenizers==0.22.2' 'huggingface-hub==0.36.2' 'accelerate==1.12.0' 'safetensors==0.7.0' 'numpy==2.2.3' 'Pillow==11.1.0'
   .venv/bin/python -m pip check
   .venv/bin/python -m pip freeze > resolved-environment.txt

   Expected result: Pinned direct versions resolve and pip check passes; exact transitive/system resolution is saved. No native execution/compatibility or fit claim.

Download the model

1. 1. Write the complete immutable component manifest

   This reader-authored manifest selects 19 exact bundle files, including all four weight files. LFS blobs use publisher SHA256; regular source files use their exact Git blob SHA1. SHA1 is Git identity, not a modern adversarial integrity guarantee; immutable official HTTPS revision remains the provenance boundary.

   CommandCopy command

   set -e
   cd "$HOME/flux2-klein4b-image"
   cat > component-manifest.json <<'MANIFEST'
   {
     "repository": "black-forest-labs/FLUX.2-klein-4B",
     "revision": "e7b7dc27f91deacad38e78976d1f2b499d76a294",
     "files": [
       {
         "file": "LICENSE.md",
         "bytes": 9584,
         "algorithm": "git-blob-sha1",
         "digest": "0afa7632c1a195ea6627a3c53e6ebf18d1be3aa5"
       },
       {
         "file": "model_index.json",
         "bytes": 446,
         "algorithm": "git-blob-sha1",
         "digest": "554a39bf1019138dd0b28809b745df8ef33657c8"
       },
       {
         "file": "scheduler/scheduler_config.json",
         "bytes": 486,
         "algorithm": "git-blob-sha1",
         "digest": "7e53605e2643440b310fa554f087641893883c5f"
       },
       {
         "file": "text_encoder/config.json",
         "bytes": 1536,
         "algorithm": "git-blob-sha1",
         "digest": "86f951d170cfd1fdc1c8a57a041e05c27b7b789b"
       },
       {
         "file": "text_encoder/generation_config.json",
         "bytes": 214,
         "algorithm": "git-blob-sha1",
         "digest": "f5af93d0c7764fa38e3b1cd3aaba1b6307246d3b"
       },
       {
         "file": "text_encoder/model-00001-of-00002.safetensors",
         "bytes": 4967215360,
         "algorithm": "sha256",
         "digest": "8c0506e7f4936fa7e26183a4fd8da4e2bdbc5990ba64ae441f965d51228f36ea"
       },
       {
         "file": "text_encoder/model-00002-of-00002.safetensors",
         "bytes": 3077766632,
         "algorithm": "sha256",
         "digest": "82f2bd839378541b0557bfabaf37c7d3d637071fdcb73302dedd7cf61162ce07"
       },
       {
         "file": "text_encoder/model.safetensors.index.json",
         "bytes": 32855,
         "algorithm": "git-blob-sha1",
         "digest": "b65d8063815c8679d85bb919337b36fae8de75cf"
       },
       {
         "file": "tokenizer/added_tokens.json",
         "bytes": 707,
         "algorithm": "git-blob-sha1",
         "digest": "b54f9135e44c1e81047e8d05cb027af8bc039eed"
       },
       {
         "file": "tokenizer/chat_template.jinja",
         "bytes": 4168,
         "algorithm": "git-blob-sha1",
         "digest": "01be9b307daa2d425f7c168c9fb145a286e0afb4"
       },
       {
         "file": "tokenizer/merges.txt",
         "bytes": 1671853,
         "algorithm": "git-blob-sha1",
         "digest": "31349551d90c7606f325fe0f11bbb8bd5fa0d7c7"
       },
       {
         "file": "tokenizer/special_tokens_map.json",
         "bytes": 613,
         "algorithm": "git-blob-sha1",
         "digest": "ac23c0aaa2434523c494330aeb79c58395378103"
       },
       {
         "file": "tokenizer/tokenizer.json",
         "bytes": 11422654,
         "algorithm": "sha256",
         "digest": "aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4"
       },
       {
         "file": "tokenizer/tokenizer_config.json",
         "bytes": 5404,
         "algorithm": "git-blob-sha1",
         "digest": "ddaf69808214a44fdd26d3785b66c1367c78277a"
       },
       {
         "file": "tokenizer/vocab.json",
         "bytes": 2776833,
         "algorithm": "git-blob-sha1",
         "digest": "4783fe10ac3adce15ac8f358ef5462739852c569"
       },
       {
         "file": "transformer/config.json",
         "bytes": 541,
         "algorithm": "git-blob-sha1",
         "digest": "f21592ae2bc9f43052e78785bde7414f219ae369"
       },
       {
         "file": "transformer/diffusion_pytorch_model.safetensors",
         "bytes": 7751109744,
         "algorithm": "sha256",
         "digest": "9f29f9edcfdae452a653ffb51a534ca4decd389952c225724ff3b94042612a6e"
       },
       {
         "file": "vae/config.json",
         "bytes": 821,
         "algorithm": "git-blob-sha1",
         "digest": "c3f38eb4c188a96a462519159356bd0fcc28cd14"
       },
       {
         "file": "vae/diffusion_pytorch_model.safetensors",
         "bytes": 168120878,
         "algorithm": "sha256",
         "digest": "ca70d2202afe6415bdbcb8793ba8cd99fd159cfe6192381504d6c4d3036e0f04"
       }
     ]
   }
   MANIFEST
   sha256sum --check <<'SHA256'
   f03107bcd53ab215deeee44072681f15658a94a755ae242bb42741dfca6405e3  component-manifest.json
   SHA256

   Expected result: Exact complete manifest verifies before any reader weight retrieval.

2. 2. Retrieve and verify all native components

   Reader-only downloads from one exact ungated publisher revision. Do not fetch the standalone transformer-only file or float revisions. Stop on any size/hash mismatch; no model loading occurs in this retrieval script.

   CommandCopy command

   set -e
   cd "$HOME/flux2-klein4b-image"
   .venv/bin/python - <<'PYDOWNLOAD'
   import hashlib,json,pathlib,shutil
   from huggingface_hub import hf_hub_download
   m=json.load(open('component-manifest.json'))
   for row in m['files']:
    src=hf_hub_download(m['repository'],row['file'],revision=m['revision'])
    dst=pathlib.Path('model')/row['file'];dst.parent.mkdir(parents=True,exist_ok=True);shutil.copyfile(src,dst)
    assert dst.stat().st_size==row['bytes']
    h=hashlib.sha256() if row['algorithm']=='sha256' else hashlib.sha1()
    if row['algorithm']=='git-blob-sha1':h.update(('blob '+str(row['bytes'])+'\0').encode())
    with dst.open('rb') as f:
     for chunk in iter(lambda:f.read(1024*1024),b''):h.update(chunk)
    assert h.hexdigest()==row['digest'],row['file']
   print('All19 components verified; model loading has not run.')
   PYDOWNLOAD

   Expected result: All 19 exact files verify. Download cache duplicates disk usage;80 GiB is an allocation, not a measured peak. Offline loading follows only after complete verification.

Configure

1. 1. Create a stable private service key

   Generate only when absent; preserve existing access. Do not print or share the local key.

   CommandCopy command

   set -e
   cd "$HOME/flux2-klein4b-image"
   umask 077
   if [ ! -e .api-key ]; then openssl rand -hex 32 > .api-key; fi
   chmod 600 .api-key
   test -s .api-key || { echo "Restore empty key before continuing"; exit 1; }

   Expected result: Private64-character lowercase hexadecimal key preserved or created.

2. 2. Create the native image harness

   Reader-authored wrapper, not publisher server. It rechecks every component before explicit native local_files_only safetensors component loading. Reject any missing, unexpected, mismatched or error diagnostic before assembling the native pipeline. All modules use CPU float32; fixed prompt is chat-formatted with thinking disabled and checked before native 512 truncation. No full model binary was fetched or executed by the author.

   CommandCopy command

   set -e
   cd "$HOME/flux2-klein4b-image"
   cat > flux2_image.py <<'PYSERVER'
   """Reader-authored educational fixed-input FLUX.2 Klein4B harness; not publisher server."""
   import hashlib
   import hmac
   import json
   import math
   import pathlib
   from http.server import BaseHTTPRequestHandler, HTTPServer
   
   ROOT = pathlib.Path(__file__).resolve().parent
   PORT = 8995
   
   
   PROMPT = 'A simple red cube beside a blue sphere on a plain white background, clean studio illustration.'
   
   
   def validate_image(result):
       import base64
       import struct
       png, count = result
       if type(count) is not int or not 0 < count <= 512:
           raise ValueError('prompt_budget')
       if not isinstance(png, bytes) or not 57 <= len(png) <= 8*1024*1024:
           raise ValueError('png_bound')
       if png[:8] != b'\x89PNG\r\n\x1a\n' or png[12:16] != b'IHDR':
           raise ValueError('png_header')
       if struct.unpack('>II', png[16:24]) != (512, 512) or png[-8:-4] != b'IEND':
           raise ValueError('png_shape_or_end')
       return {'fixture': 'author-t2i-v1', 'prompt': PROMPT,
               'prompt_tokens_before_padding': count, 'padded_prompt_positions': 512,
               'image_shape': [512, 512, 3], 'configured_image_latent_positions': 1024,
               'denoising_steps': 4, 'seed': 0, 'guidance_scale': 1.0,
               'mime_type': 'image/png', 'png_base64': base64.b64encode(png).decode('ascii'),
               'generated_text_tokens': None}
   
   
   def make_encoder():
       # Reader-run only. Fake tests do not import Torch, Diffusers or Pillow.
       import os
       os.environ['HF_HUB_OFFLINE'] = '1'
       os.environ['TRANSFORMERS_OFFLINE'] = '1'
       import torch
       from diffusers import (Flux2KleinPipeline, Flux2Transformer2DModel,
                              AutoencoderKLFlux2, FlowMatchEulerDiscreteScheduler)
       from transformers import Qwen3ForCausalLM, Qwen2TokenizerFast
       manifest = json.loads((ROOT / 'component-manifest.json').read_text())
       model_root = ROOT / 'model'
       for row in manifest['files']:
           file = model_root / row['file']
           if file.stat().st_size != row['bytes']:
               raise RuntimeError('component_size')
           hasher = hashlib.sha256() if row['algorithm'] == 'sha256' else hashlib.sha1()
           if row['algorithm'] == 'git-blob-sha1':
               hasher.update(('blob '+str(row['bytes'])+'\0').encode())
           with file.open('rb') as stream:
               for chunk in iter(lambda: stream.read(1024*1024), b''):
                   hasher.update(chunk)
           if hasher.hexdigest() != row['digest']:
               raise RuntimeError('component_hash')
       torch.set_num_threads(8)
       def load_checked(cls, subfolder):
           component, info = cls.from_pretrained(
               str(model_root / subfolder), torch_dtype=torch.float32,
               local_files_only=True, use_safetensors=True, output_loading_info=True)
           for key in ['missing_keys', 'unexpected_keys', 'mismatched_keys', 'error_msgs']:
               if info.get(key):
                   raise RuntimeError('component_loading_diagnostics_'+subfolder+'_'+key)
           return component.float().cpu().eval()
       transformer = load_checked(Flux2Transformer2DModel, 'transformer')
       text_encoder = load_checked(Qwen3ForCausalLM, 'text_encoder')
       vae = load_checked(AutoencoderKLFlux2, 'vae')
       tokenizer = Qwen2TokenizerFast.from_pretrained(str(model_root / 'tokenizer'), local_files_only=True)
       scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
           str(model_root / 'scheduler'), local_files_only=True)
       pipe = Flux2KleinPipeline(transformer=transformer, text_encoder=text_encoder,
                                 vae=vae, tokenizer=tokenizer, scheduler=scheduler,
                                 is_distilled=True)
       pipe = pipe.to('cpu')
       pipe.set_progress_bar_config(disable=True)
       if not json.loads((model_root / 'model_index.json').read_text())['is_distilled'] or not pipe.config.is_distilled:
           raise RuntimeError('wrong_variant')
       formatted = pipe.tokenizer.apply_chat_template(
           [{'role': 'user', 'content': PROMPT}], tokenize=False,
           add_generation_prompt=True, enable_thinking=False)
       token_count = len(pipe.tokenizer(formatted, truncation=False)['input_ids'])
       if not 0 < token_count <= 512:
           raise RuntimeError('prompt_would_truncate')
   
       def encode():
           import io
           import numpy as np
           from PIL import Image
           def check_latents(pipeline, step, timestep, values):
               if not torch.isfinite(values['latents']).all():
                   raise RuntimeError('nonfinite_latents')
               return values
           with torch.inference_mode():
               image = pipe(prompt=PROMPT, height=512, width=512, num_inference_steps=4,
                            guidance_scale=1.0, max_sequence_length=512,
                            num_images_per_prompt=1, output_type='np',
                            callback_on_step_end=check_latents,
                            callback_on_step_end_tensor_inputs=['latents'],
                            generator=torch.Generator(device='cpu').manual_seed(0)).images[0]
           if image.shape != (512, 512, 3) or not np.isfinite(image).all():
               raise RuntimeError('image_shape_or_finiteness')
           if image.min() < 0 or image.max() > 1:
               raise RuntimeError('image_range')
           image = Image.fromarray((image*255).round().astype('uint8'), mode='RGB')
           buf = io.BytesIO()
           image.save(buf, format='PNG')
           return buf.getvalue(), token_count
       return encode
   
   
   PYSERVER

   Expected result: First complete native loader and PNG-validation fragment written.

3. 3. Complete and verify private transport

   Append fixed-input private serial HTTP transport. No model code runs in syntax/hash checks.

   CommandCopy command

   set -e
   cd "$HOME/flux2-klein4b-image"
   cat >> flux2_image.py <<'PYSERVER'
   def create_server(encoder, key, port=PORT):
       if not isinstance(key, str) or len(key) != 64 or any(c not in '0123456789abcdef' for c in key):
           raise ValueError('key_format')
   
       class Handler(BaseHTTPRequestHandler):
           def setup(self):
               super().setup()
               self.connection.settimeout(5)
   
           def log_message(self, *_):
               pass  # Do not log Bearer keys or request bodies.
   
           def reply(self, status, obj):
               data = json.dumps(obj, allow_nan=False, separators=(',', ':')).encode()
               self.send_response(status)
               self.send_header('Content-Type', 'application/json')
               self.send_header('Content-Length', str(len(data)))
               self.send_header('Connection', 'close')
               self.end_headers()
               self.wfile.write(data)
               self.close_connection = True
   
           def authorized(self):
               values = self.headers.get_all('Authorization', [])
               return len(values) == 1 and hmac.compare_digest(values[0].encode(), ('Bearer ' + key).encode())
   
           def do_GET(self):
               if not self.authorized():
                   return self.reply(401, {'error': 'unauthorized'})
               if self.path != '/health':
                   return self.reply(404, {'error': 'not_found'})
               return self.reply(200, {'ready': True, 'padded_prompt_positions': 512, 'configured_image_latent_positions': 1024,
                                       'generated_text_tokens': None})
   
           def do_POST(self):
               if not self.authorized():
                   return self.reply(401, {'error': 'unauthorized'})
               if self.path != '/v1/fixed-image':
                   return self.reply(404, {'error': 'not_found'})
               if self.headers.get('Transfer-Encoding') is not None:
                   return self.reply(400, {'error': 'transfer_encoding_unsupported'})
               lengths = self.headers.get_all('Content-Length', [])
               if len(lengths) != 1:
                   return self.reply(411, {'error': 'content_length_required'})
               raw = lengths[0]
               if not raw or len(raw) > 4 or any(c not in '0123456789' for c in raw):
                   return self.reply(413, {'error': 'invalid_or_excessive_length'})
               n = int(raw)
               if n < 1 or n > 256:
                   return self.reply(413, {'error': 'body_limit'})
               if self.headers.get_content_type() != 'application/json':
                   return self.reply(415, {'error': 'json_required'})
               try:
                   body = self.rfile.read(n)
               except (TimeoutError, OSError):
                   return self.reply(408, {'error': 'body_timeout'})
               if len(body) != n:
                   return self.reply(400, {'error': 'incomplete_body'})
               try:
                   def unique(pairs):
                       obj = {}
                       for k, v in pairs:
                           if k in obj:
                               raise ValueError('duplicate_key')
                           obj[k] = v
                       return obj
                   request = json.loads(body.decode('utf-8'), object_pairs_hook=unique,
                                        parse_constant=lambda _: (_ for _ in ()).throw(ValueError('constant')))
               except (ValueError, UnicodeError, RecursionError):
                   return self.reply(400, {'error': 'invalid_json'})
               if type(request) is not dict or request != {'fixture': 'author-t2i-v1'}:
                   return self.reply(400, {'error': 'fixed_fixture_required'})
               try:
                   result = validate_image(encoder())
               except Exception:
                   return self.reply(500, {'error': 'pipeline_or_png_failure'})
               return self.reply(200, result)
   
       return HTTPServer(('127.0.0.1', port), Handler)
   
   
   if __name__ == '__main__':
       secret = (ROOT / '.api-key').read_text().strip()
       encoder = make_encoder()
       with create_server(encoder, secret) as server:
           server.serve_forever()
   PYSERVER
   .venv/bin/python -m py_compile flux2_image.py
   sha256sum --check <<'SHA256'
   d4e624292a3cec1f07856e464c0020b3d2db6591555e04ffe673ac3f2dffaefb  flux2_image.py
   SHA256

   Expected result: Exact wrapper bytes and Python syntax verify without model loading.

Start the model

1. 1. Start the complete CPU image experiment

   Foreground serial endpoint. Full pipeline load and first image may be slow or exceed RAM; no fit/speed is promised. Explicit native loaders select all model_index components, reject loading diagnostics and assemble the native pipeline with offline safetensors and CPU float32; no VAE/text-encoder substitution. Each denoising-step latent tensor and final NumPy image must be finite before PNG conversion.

   CommandCopy command

   set -e
   cd "$HOME/flux2-klein4b-image"
   HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 .venv/bin/python flux2_image.py

   Expected result: After actual reader-side component verification and native load, loopback 127.0.0.1:8995 serves. Stop on component/import/load/template errors.

Verify the setup

1. 1. Validate the PNG and inspect the generated image

   Run in a second server terminal. Check private health, missing-key 401, fixed response metadata and a valid decoded 512 RGB PNG. Inspect image.png yourself for the red cube/blue sphere on white; valid dimensions do not establish semantic fidelity, safety or quality. No generated example or success claim is supplied.

   CommandCopy command

   set -e
   cd "$HOME/flux2-klein4b-image"
   FLUX_KEY="$(cat .api-key)"
   curl --fail http://127.0.0.1:8995/health -H "Authorization: Bearer $FLUX_KEY"
   curl --fail http://127.0.0.1:8995/v1/fixed-image -H "Authorization: Bearer $FLUX_KEY" -H 'Content-Type: application/json' -d '{"fixture":"author-t2i-v1"}' -o image.json
   curl -s -o /dev/null -w '%{http_code}\n' http://127.0.0.1:8995/health
   unset FLUX_KEY
   .venv/bin/python - <<'PYVERIFY'
   import base64,io,json
   from PIL import Image
   d=json.load(open('image.json'));assert d['generated_text_tokens'] is None
   assert d['image_shape']==[512,512,3] and d['padded_prompt_positions']==512
   assert 0<d['prompt_tokens_before_padding']<=512 and d['configured_image_latent_positions']==1024
   assert d['denoising_steps']==4 and d['seed']==0
   b=base64.b64decode(d['png_base64'],validate=True);assert len(b)<=8*1024*1024
   with Image.open(io.BytesIO(b)) as image:
    assert image.format=='PNG' and image.size==(512,512) and image.mode=='RGB';image.verify()
   open('image.png','wb').write(b)
   print('Inspect image.png; PNG validity is not semantic quality proof.')
   PYVERIFY

   Expected result: Authenticated health/generation return structured JSON; missing auth 401. Exact PNG metadata/Pillow verify passes only after reader execution. Inspect and record actual visual defects or loading failures.

Performance evidence

No compatible measurement available. Documented support and estimated memory do not establish response speed.

Limits and troubleshooting

• Complete official 4B distilled bundle is selected, not transformer-only 7.75 GB weights or9B/base/dev. Main tensor file metadata plus all required configs/processors are pinned; no model binary was downloaded.
• 512 configured prompt positions include the native user template with thinking disabled. Actual pre-padding prompt count is checked at reader execution; too-long prompts stop before native silent truncation. This is not 40960 theoretical Qwen3 context or 1024 image latent positions.
• Native four-step distilled image synthesis returns PNG pixels, not autoregressive output tokens. Guidance_scale 1.0 avoids CFG; no chat completion/reasoning output or text-generation benchmark is claimed.
• Full CPU float32 through native .to(cpu) is source-qualified, not installed/executed compatibility proof. 64 GiB RAM, 80 GiB disk/eight threads are planning allocations; fit, peak memory, latency and image quality remain unknown.
• Original fixed geometric prompt is not a third-party image fixture. Manual inspection remains necessary; fixed seed does not guarantee equal bytes across hardware/library versions. No downstream safety classifier or commercial suitability is claimed.
• Reader-authored serial HTTPServer has private key/loopback, bounded strict fixed JSON and PNG-size/shape guards, but no inference deadline, multiuser robustness or public production-service guarantee.
• All components are verified before offline local loading; no remote-code trust or unpinned auxiliary download is selected. Publisher Git blob SHA1 is identified accurately; weight SHA256s are metadata until reader actual download.
• Fake PNG protocol tests prove only request/auth/output boundary handling, not actual native tensor finiteness, denoising correctness, prompt adherence, image safety or runtime fit. Native callbacks and pre-PNG guards require finite tensors/arrays at reader execution; they have not run on model weights here.
• Image editing, multiple references, other resolutions, base variants, alternate supported GPU/quantized/offload backends and other useful modes remain open; this pair does not finalize family coverage.

Component/hash/version or offline load error

Stop and restore exact 19-file manifest and publisher revision. Do not float to main, omit text encoder/VAE or substitute standalone transformer-only weights. Keep safetensors/local_files_only; stop on unresolved dependency errors.

Prompt budget or wrong variant error

Use exact fixed prompt/tokenizer/template and distilled model_index. The pre-padding formatted prompt must fit 512 positions. Do not hide truncation, switch model variants or treat padding count as actual semantic prompt length.

401 or rejected body

Use preserved matching private 64-character key and exact fixture selector. Arbitrary prompts/images/URLs/files or extra fields are rejected.

Slow, out-of-memory, black or inaccurate image

Record actual load/generation failure and inspect PNG. 64 GiB is not measured fit and valid dimensions do not establish semantic quality. Do not claim a smaller/partial pipeline is equivalent; stop the process if resource use is unacceptable.

Stop the model

1. 1. Stop the private image service

   Press Ctrl+C in its foreground terminal and unset client key. Close optional SSH tunnel; stopping Python does not stop VM billing.

   CommandCopy command

   unset FLUX_KEY

   Expected result: Pipeline process and endpoint stop; generated files remain until explicitly removed.

Uninstall

1. 1. Remove dedicated experiment and selected model cache

   Preserve desired outputs/audit/license records first. Confirm the dedicated directory before deletion. The model-specific download cache is separate; remove only this exact cache if no other experiment needs it. Shared Ubuntu packages remain.

   CommandCopy command

   rm -r -- "$HOME/flux2-klein4b-image"
   rm -r -- "$HOME/.cache/huggingface/hub/models--black-forest-labs--FLUX.2-klein-4B"

   Expected result: Dedicated bundle/venv/wrapper/key/outputs and this selected model cache are removed if present.

Shut down cloud resources

1. 1. End cloud billing resources

   In EC2 us-east-1, stop for later reuse or terminate this exact VM. Review Delete on termination for its root volume. Delete unwanted detached EBS volumes, snapshots and reserved/Elastic/public IP resources. Stopped VM storage and retained resources may still bill. Check Bills/Cost Explorer later for final usage.

   Expected result: Exact instance is stopped/terminated as intended; no unwanted attached/retained billed resources remain. Instance compute stops only after its lifecycle transition, not when the Python service stops.