MiniCPM

Exact model and requirements

Artifact:

OpenBMB / Native MiniCPM4-0.5B · checkpoint openbmb/MiniCPM4-0.5B

Publisher revision:

5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630

Artifact digest:

sha256:ef74f4de1d1c580dd64e425d11f29881b9fd39f9be16a485150f727d4583181c

Download size:

0.81 GiB

Memory requirement (unknown):

Unknown — Native full FP32 process/loading/activations/OS peak is unmeasured.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host and one sequence. All native language model computation is on CPU; no GPU or disk offload.

KV cache:

Not allocated by this recipe: native use_cache=False; eager recomputation on CPU.

Workload:

4,096 token context; Unknown; Unknown; concurrency 1; batch 1; cache cold; repetitions Unknown.

Tokenizer:

Native BPE tokenizer JSON and exact MiniCPM4 chat template · Native checkpoint pin; tokenizers 0.22.2

Sampling / reasoning:

Publisher card uses temperature 0.7 and top_p 0.7. This walkthrough uses those values, top_k 0 (author choice), repetition penalty 1.0 and one beam, with native EOS IDs 2 and 73440 and pad ID 2. Output sampling is nondeterministic; no exact answer is promised. This exact small MiniCPM4 text checkpoint uses its native ordinary user/assistant template. No hybrid reasoning toggle is introduced. The native configuration allows 32,768 positions; 4,096 is an author-selected bounded walkthrough, not a claim of full-family or long-context support. Render the native template, then tokenize once with its BOS postprocessor. Disable native cache explicitly.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

The exact native model card declares Apache 2.0 and links the publisher repository license. Preserve the license, applicable notices, modification notices and patent conditions. No inference quality, suitability, endorsement or legal clearance is claimed. Review the archived actual publisher license before use.

• model/config.json · 0 GiB · revision 5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630
• model/generation_config.json · 0 GiB · revision 5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630
• model/configuration_minicpm.py · 0 GiB · revision 5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630
• model/modeling_minicpm.py · 0 GiB · revision 5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630
• model/tokenizer.json · 0.01 GiB · revision 5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630
• model/tokenizer_config.json · 0 GiB · revision 5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630
• model/special_tokens_map.json · 0 GiB · revision 5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630
• model/added_tokens.json · 0 GiB · revision 5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630

• The exact serialized native weight file is 867,772,648 bytes. FP32 expansion, loading transients, tokenizer/runtime and OS buffers require additional memory; file size is not RAM fit.
• 32 GiB RAM, 40 GiB disk and four CPU threads/vCPUs are planning allocations. Eager recomputation with cache disabled may be slow; no CPU timing or complete-process fit was measured.

Install the runtime

1. 1. Prepare the native CPU environment

   Install the exact CPU dependency set in the dedicated virtual environment. Package compatibility is source-qualified; installation was not executed by the author.

   CommandCopy command

   set -e
   sudo apt-get update
   sudo apt-get install -y python3.12 python3.12-venv curl ca-certificates openssl
   mkdir -p "$HOME/minicpm4-05b/model"
   cd "$HOME/minicpm4-05b"
   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'
   .venv/bin/python -m pip install 'transformers==4.57.6' 'accelerate==1.10.1' 'safetensors==0.5.3' 'tokenizers==0.22.2' 'huggingface-hub==0.36.0' 'numpy==2.2.3'
   .venv/bin/python -m pip check
   .venv/bin/python -m pip freeze > resolved-environment.txt

   Expected result: Dedicated virtual environment, exact CPU packages and a passing dependency check; stop on any package or version error.

Download the model

1. 1. Download the exact native checkpoint and source files

   Download all nine pinned model, tokenizer, configuration and custom-code assets. Streaming size/checksum checks fail fast. The native tokenizer JSON is used directly; no tokenizer conversion or extra vocabulary is created. Read the custom Python before continuing.

   CommandCopy command

   set -eu
   cd "$HOME/minicpm4-05b"
   mkdir -p model
   curl -fL --retry 0 'https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/config.json' -o 'model/config.json'
   curl -fL --retry 0 'https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/generation_config.json' -o 'model/generation_config.json'
   curl -fL --retry 0 'https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/configuration_minicpm.py' -o 'model/configuration_minicpm.py'
   curl -fL --retry 0 'https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/modeling_minicpm.py' -o 'model/modeling_minicpm.py'
   curl -fL --retry 0 'https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/tokenizer.json' -o 'model/tokenizer.json'
   curl -fL --retry 0 'https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/tokenizer_config.json' -o 'model/tokenizer_config.json'
   curl -fL --retry 0 'https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/special_tokens_map.json' -o 'model/special_tokens_map.json'
   curl -fL --retry 0 'https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/added_tokens.json' -o 'model/added_tokens.json'
   curl -fL --retry 0 'https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/model.safetensors' -o 'model/model.safetensors'
   cat > asset-manifest.json <<'JSONMANIFEST'
   {
     "model/config.json": {
       "bytes": 2403,
       "sha256": "610c5b519420a3633ec4950c15a26d83ad0e9be35ef9d00a9e3d2b621b325f72",
       "url": "https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/config.json"
     },
     "model/generation_config.json": {
       "bytes": 182,
       "sha256": "83595cf3b045d7ed16593ae399b0e3c13fe45640a4bec3a1cd371b057a6a98b7",
       "url": "https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/generation_config.json"
     },
     "model/configuration_minicpm.py": {
       "bytes": 9660,
       "sha256": "1e289912fdc0608ccad3983f4ff8b381ec8f161f45c85be4d694f92f17596b0a",
       "url": "https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/configuration_minicpm.py"
     },
     "model/modeling_minicpm.py": {
       "bytes": 69994,
       "sha256": "c646022c44d0e1cc1d149fb80422a5d5f1ba6de66dd0f525f2be1e18d99994bf",
       "url": "https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/modeling_minicpm.py"
     },
     "model/tokenizer.json": {
       "bytes": 6701371,
       "sha256": "adf7208af154a5ca065d2eda4e5419e02aac58c2c00627874748b75ec6769094",
       "url": "https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/tokenizer.json"
     },
     "model/tokenizer_config.json": {
       "bytes": 2835,
       "sha256": "8bef75d59004eb0e62db89e97090d4736b14fd6c4a0282370c28833be1e83051",
       "url": "https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/tokenizer_config.json"
     },
     "model/special_tokens_map.json": {
       "bytes": 630,
       "sha256": "9ec9d6f07cb0e436598bf29b4d97c6355434c08fdbe9c264e6926907f3751c7e",
       "url": "https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/special_tokens_map.json"
     },
     "model/added_tokens.json": {
       "bytes": 216,
       "sha256": "4760fcbf90bc193f33827ffe02f2e7ba1af1ec43644cc02ac22fdd611f6cca15",
       "url": "https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/added_tokens.json"
     },
     "model/model.safetensors": {
       "bytes": 867772648,
       "sha256": "ef74f4de1d1c580dd64e425d11f29881b9fd39f9be16a485150f727d4583181c",
       "url": "https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/model.safetensors"
     }
   }
   JSONMANIFEST
   .venv/bin/python - <<'PYHASH'
   import pathlib,json,hashlib
   for name,info in json.load(open('asset-manifest.json')).items():
    p=pathlib.Path(name);digest=hashlib.sha256()
    with p.open('rb') as stream:
     for chunk in iter(lambda:stream.read(1024*1024),b''):digest.update(chunk)
    assert p.stat().st_size==info['bytes'] and digest.hexdigest()==info['sha256'],name
   print('All nine native asset checks passed.')
   PYHASH

   Expected result: All nine exact asset checks pass. Stop on an access error, byte-size mismatch or checksum mismatch.

Configure

1. 1. Create a private key

   Keep the key in this private directory. Never print it or share it. Requests and health checks require the same key.

   CommandCopy command

   set -eu
   cd "$HOME/minicpm4-05b"
   umask 077
   .venv/bin/python - <<'PYKEY'
   import pathlib,secrets
   p=pathlib.Path('.api-key')
   with p.open('x') as f:f.write(secrets.token_hex(32)+'\n')
   PYKEY

   Expected result: Private key file created with restrictive permissions; no key printed.

2. 2. Write the checked native CPU loader

   The loader streams all asset checksums, rejects all native loading diagnostics and uses full FP32 eager CPU execution. Trust is limited to the exact reviewed local MiniCPM source. Cache is explicitly disabled at generation.

   CommandCopy command

   cd "$HOME/minicpm4-05b"
   cat > reader-chat-server.py <<'PYSERVER1'
   import pathlib,json,hashlib,hmac
   from http.server import BaseHTTPRequestHandler,HTTPServer
   ROOT=pathlib.Path(__file__).resolve().parent
   PORT=8994
   
   def valid_prompt(prompt):
       return type(prompt) is str and bool(prompt.strip()) and len(prompt.encode('utf-8'))<=4096 and not any(x in prompt for x in ['<|','<s>','</s>'])
   
   def validate_completion(result):
       n,ids,answer=result
       if type(n) is not int or not 1<=n<=3584:raise ValueError('prompt_budget')
       if not isinstance(ids,list) or not 2<=len(ids)<512 or any(type(x) is not int or x<0 or x>=73448 for x in ids):raise ValueError('output_budget_or_ids')
       if ids[-1] not in (2,73440) or any(x in (2,73440) for x in ids[:-1]):raise ValueError('native_eos_required')
       if type(answer) is not str or not answer.strip() or len(answer.encode('utf-8'))>16384:raise ValueError('empty_or_excessive_answer')
       return {'answer':answer,'prompt_tokens':n,'generated_tokens_including_eos':len(ids),'complete_by_native_eos':True,'quality_measured':False}
   
   def validate_loading_info(info):
       fields={'missing_keys','unexpected_keys','mismatched_keys','error_msgs'}
       if type(info) is not dict or set(info)!=fields or any(type(info[k]) is not list or info[k] for k in fields):raise RuntimeError('incomplete_native_weight_load')
   
   def make_encoder():
       import os
       os.environ['HF_HUB_OFFLINE']='1';os.environ['TRANSFORMERS_OFFLINE']='1'
       import torch
       from transformers import AutoModelForCausalLM,PreTrainedTokenizerFast
       torch.set_num_threads(4)
       manifest=json.loads((ROOT/'asset-manifest.json').read_text())
       for name,info in manifest.items():
           file=ROOT/name;digest=hashlib.sha256()
           with file.open('rb') as stream:
               for chunk in iter(lambda:stream.read(1024*1024),b''):digest.update(chunk)
           if file.stat().st_size!=info['bytes'] or digest.hexdigest()!=info['sha256']:raise RuntimeError('asset_mismatch:'+name)
       tokenizer=PreTrainedTokenizerFast.from_pretrained(ROOT/'model',local_files_only=True)
       if tokenizer.bos_token_id!=1 or tokenizer.eos_token_id!=73440 or len(tokenizer)!=73448:raise RuntimeError('native_tokenizer_configuration')
       model,info=AutoModelForCausalLM.from_pretrained(ROOT/'model',local_files_only=True,trust_remote_code=True,
           dtype=torch.float32,attn_implementation='eager',use_safetensors=True,output_loading_info=True,low_cpu_mem_usage=True)
       validate_loading_info(info)
       cfg=model.config
       if not (cfg.model_type=='minicpm' and cfg.hidden_size==1024 and cfg.num_hidden_layers==24 and cfg.num_attention_heads==16 and cfg.num_key_value_heads==2 and cfg.max_position_embeddings==32768):raise RuntimeError('native_architecture')
       model=model.to('cpu').float().eval()
       for param in model.parameters():param.requires_grad_(False)
       template=json.loads((ROOT/'model/tokenizer_config.json').read_text())['chat_template']
       def encode(prompt):
           text=tokenizer.apply_chat_template([{'role':'user','content':prompt}],chat_template=template,tokenize=False,add_generation_prompt=True)
           inputs=tokenizer(text,add_special_tokens=True,return_tensors='pt',truncation=False)
           n=inputs['input_ids'].shape[1]
           if n<1 or n>3584 or inputs['input_ids'][0,0].item()!=1 or (inputs['input_ids']==1).sum().item()!=1:raise ValueError('prompt_budget_or_bos')
           with torch.inference_mode():
               full=model.generate(**inputs,max_new_tokens=512,do_sample=True,temperature=0.7,top_p=0.7,top_k=0,
                   repetition_penalty=1.0,num_beams=1,eos_token_id=[2,73440],pad_token_id=2,use_cache=False)
           if full.shape[0]!=1 or not torch.equal(full[0,:n],inputs['input_ids'][0]):raise ValueError('prefix_changed')
           ids=full[0,n:].tolist()
           return n,ids,tokenizer.decode(ids,skip_special_tokens=True)
       return encode
   
   PYSERVER1

   Expected result: Native loader and protocol validation written; no model loaded yet.

3. 3. Write the private serial HTTP boundary

   Complete the server. It accepts one plain prompt with strict framing, authentication and byte/token limits; template control strings are rejected. It performs no tool action, URL fetch or file upload.

   CommandCopy command

   cd "$HOME/minicpm4-05b"
   cat >> reader-chat-server.py <<'PYSERVER2'
   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, 'context_tokens': 4096,
                                       'output_cap': 512})
   
           def do_POST(self):
               if not self.authorized():
                   return self.reply(401, {'error': 'unauthorized'})
               if self.path != '/v1/chat':
                   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 > 8192:
                   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 set(request)!={'prompt'} or not valid_prompt(request['prompt']):
                   return self.reply(400, {'error': 'bounded_plain_prompt_required'})
               try:
                   result = validate_completion(encoder(request['prompt']))
               except Exception:
                   return self.reply(500, {'error': 'generation_or_protocol_failure'})
               return self.reply(200, result)
   
       return HTTPServer(('127.0.0.1', port), Handler)
   
   
   if __name__=='__main__':
       encoder=make_encoder()
       with create_server(encoder,(ROOT/'.api-key').read_text().strip()) as server:server.serve_forever()
   PYSERVER2

   Expected result: Complete private serial server written; model work starts only in the next step.

Start the model

1. 1. Start the native model process

   Keep this foreground terminal open. Startup first validates all files and native loading diagnostics. Readiness occurs only after the actual model finishes loading; stop on any failure rather than relaxing checks.

   CommandCopy command

   cd "$HOME/minicpm4-05b"
   .venv/bin/python reader-chat-server.py

   Expected result: Foreground process remains running after a successful native load. Memory fit and startup time must be observed by the reader; no result was measured here.

Verify the setup

1. 1. Check readiness, actual text completion and authentication

   Use a second terminal on the same host. Commands use the virtual environment by absolute relative path and require no prior activation. Inspect the actual answer and record semantic errors. EOS, shape and token-budget validity are separate from quality. A client timeout does not cancel native inference.

   CommandCopy command

   set -eu
   cd "$HOME/minicpm4-05b"
   MINICPM_KEY=$(cat .api-key)
   export MINICPM_KEY
   curl -fsS --max-time 30 -H "Authorization: Bearer $MINICPM_KEY" http://127.0.0.1:8994/health -o health.json
   curl -fsS --max-time 1800 -H "Authorization: Bearer $MINICPM_KEY" -H 'Content-Type: application/json' --data '{"prompt":"Describe a quiet garden in two short sentences."}' http://127.0.0.1:8994/v1/chat -o answer.json
   .venv/bin/python - <<'PYCHECK'
   import json
   h=json.load(open('health.json'));r=json.load(open('answer.json'))
   assert h=={'ready':True,'context_tokens':4096,'output_cap':512}
   assert r['complete_by_native_eos'] is True and r['quality_measured'] is False
   assert 1<=r['prompt_tokens']<=3584 and 2<=r['generated_tokens_including_eos']<512
   assert isinstance(r['answer'],str) and r['answer'].strip()
   print(r['answer'])
   print('Protocol complete. Record whether this actual answer describes a garden and follows the requested length; this is human assessment, not a forced model result.')
   PYCHECK
   code=$(curl -sS --max-time 30 -o unauthorized.json -w '%{http_code}' http://127.0.0.1:8994/health)
   test "$code" = 401
   unset MINICPM_KEY

   Expected result: Authenticated health succeeds, an actual nonempty answer terminates at a native EOS below the cap, and unauthenticated health returns 401. No exact wording or quality is promised.

Performance evidence

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

Limits and troubleshooting

• Only this exact native MiniCPM4-0.5B text checkpoint and bounded CPU path are source-qualified. Other MiniCPM4/4.1 sizes, sparse/hybrid reasoning, quantization, GPU, long-context, tool and newer-family workflows remain open.
• The native configuration has 32,768 positions; this author selects a 4,096-token budget with 512 reserved output tokens. Tokenizer metadata contains a sentinel maximum rather than a usable model limit; it is not used as context evidence. Actual prompt/output counts remain unmeasured until execution.
• The native template has no BOS; its pinned tokenizer postprocessor adds one BOS during the single tokenization pass. The wrapper rejects duplicate BOS, template-control input, truncated prompts, partial cap-length completions and every native loading diagnostic.
• Custom native Python executes only after reader review of the exact pinned files. Source compatibility is not an executed installation or load. All model work uses CPU FP32 eager operations with cache disabled, recomputing the sequence rather than making a cache-performance claim.
• The serial HTTP boundary has authentication, loopback, body and socket limits, but no generation deadline, public TLS or multiuser resilience. Socket/client timeouts do not cancel generation. Actual answer quality, fit and speed are unknown.

Asset, custom code or load diagnostic mismatch

Stop. Restore the exact pinned download and native runtime. Do not ignore missing, unexpected or mismatched keys, substitute a different model, enable network code fallback or relax checks.

Empty answer, no EOS, output cap or prompt overflow

Record the actual failure. Do not silently truncate inputs or accept partial answers. Keep the selected total budget and requalify any changed configuration.

Incorrect or unhelpful answer

Record the actual answer and human assessment. Protocol validity does not establish task quality or a benchmark score.

OOM, slow CPU or dependency error

Stop and record the actual environment and error. Planning capacity is not proven fit. Eager cache-disabled inference may be slow; any changed runtime or precision requires requalification.

401 or connection refused

Check foreground readiness, private key and loopback or SSH boundary. Keep the service port closed publicly and never print the key.

Stop the model

1. 1. Stop the private model process

   Press Ctrl+C in the foreground terminal and unset the key in each client terminal. Confirm the loopback endpoint no longer responds.

   CommandCopy command

   unset MINICPM_KEY

   Expected result: Private Python service stops and the model process exits.

Uninstall

1. 1. Remove the dedicated workspace

   Back up wanted results and license records, verify the exact directory and remove only this workspace. Shared Ubuntu packages remain installed.

   CommandCopy command

   rm -r -- "$HOME/minicpm4-05b"

   Expected result: Dedicated model, custom code, tokenizer, virtual environment, reader wrapper, outputs and private key removed.

Exact model and requirements

Artifact:

OpenBMB / Native MiniCPM4-0.5B · checkpoint openbmb/MiniCPM4-0.5B

Publisher revision:

5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630

Artifact digest:

sha256:ef74f4de1d1c580dd64e425d11f29881b9fd39f9be16a485150f727d4583181c

Download size:

0.81 GiB

Memory requirement (unknown):

Unknown — Native full FP32 process/loading/activations/OS peak is unmeasured.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host and one sequence. All native language model computation is on CPU; no GPU or disk offload.

KV cache:

Not allocated by this recipe: native use_cache=False; eager recomputation on CPU.

Workload:

4,096 token context; Unknown; Unknown; concurrency 1; batch 1; cache cold; repetitions Unknown.

Tokenizer:

Native BPE tokenizer JSON and exact MiniCPM4 chat template · Native checkpoint pin; tokenizers 0.22.2

Sampling / reasoning:

Publisher card uses temperature 0.7 and top_p 0.7. This walkthrough uses those values, top_k 0 (author choice), repetition penalty 1.0 and one beam, with native EOS IDs 2 and 73440 and pad ID 2. Output sampling is nondeterministic; no exact answer is promised. This exact small MiniCPM4 text checkpoint uses its native ordinary user/assistant template. No hybrid reasoning toggle is introduced. The native configuration allows 32,768 positions; 4,096 is an author-selected bounded walkthrough, not a claim of full-family or long-context support. Render the native template, then tokenize once with its BOS postprocessor. Disable native cache explicitly.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

The exact native model card declares Apache 2.0 and links the publisher repository license. Preserve the license, applicable notices, modification notices and patent conditions. No inference quality, suitability, endorsement or legal clearance is claimed. Review the archived actual publisher license before use.

• model/config.json · 0 GiB · revision 5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630
• model/generation_config.json · 0 GiB · revision 5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630
• model/configuration_minicpm.py · 0 GiB · revision 5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630
• model/modeling_minicpm.py · 0 GiB · revision 5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630
• model/tokenizer.json · 0.01 GiB · revision 5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630
• model/tokenizer_config.json · 0 GiB · revision 5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630
• model/special_tokens_map.json · 0 GiB · revision 5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630
• model/added_tokens.json · 0 GiB · revision 5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630

• The exact serialized native weight file is 867,772,648 bytes. FP32 expansion, loading transients, tokenizer/runtime and OS buffers require additional memory; file size is not RAM fit.
• 32 GiB RAM, 40 GiB disk and four CPU threads/vCPUs are planning allocations. Eager recomputation with cache disabled may be slow; no CPU timing or complete-process fit was measured.

Install the runtime

1. 1. Prepare the native CPU environment

   Install the exact CPU dependency set in the dedicated virtual environment. Package compatibility is source-qualified; installation was not executed by the author.

   CommandCopy command

   set -e
   sudo apt-get update
   sudo apt-get install -y python3.12 python3.12-venv curl ca-certificates openssl
   mkdir -p "$HOME/minicpm4-05b/model"
   cd "$HOME/minicpm4-05b"
   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'
   .venv/bin/python -m pip install 'transformers==4.57.6' 'accelerate==1.10.1' 'safetensors==0.5.3' 'tokenizers==0.22.2' 'huggingface-hub==0.36.0' 'numpy==2.2.3'
   .venv/bin/python -m pip check
   .venv/bin/python -m pip freeze > resolved-environment.txt

   Expected result: Dedicated virtual environment, exact CPU packages and a passing dependency check; stop on any package or version error.

Download the model

1. 1. Download the exact native checkpoint and source files

   Download all nine pinned model, tokenizer, configuration and custom-code assets. Streaming size/checksum checks fail fast. The native tokenizer JSON is used directly; no tokenizer conversion or extra vocabulary is created. Read the custom Python before continuing.

   CommandCopy command

   set -eu
   cd "$HOME/minicpm4-05b"
   mkdir -p model
   curl -fL --retry 0 'https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/config.json' -o 'model/config.json'
   curl -fL --retry 0 'https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/generation_config.json' -o 'model/generation_config.json'
   curl -fL --retry 0 'https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/configuration_minicpm.py' -o 'model/configuration_minicpm.py'
   curl -fL --retry 0 'https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/modeling_minicpm.py' -o 'model/modeling_minicpm.py'
   curl -fL --retry 0 'https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/tokenizer.json' -o 'model/tokenizer.json'
   curl -fL --retry 0 'https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/tokenizer_config.json' -o 'model/tokenizer_config.json'
   curl -fL --retry 0 'https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/special_tokens_map.json' -o 'model/special_tokens_map.json'
   curl -fL --retry 0 'https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/added_tokens.json' -o 'model/added_tokens.json'
   curl -fL --retry 0 'https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/model.safetensors' -o 'model/model.safetensors'
   cat > asset-manifest.json <<'JSONMANIFEST'
   {
     "model/config.json": {
       "bytes": 2403,
       "sha256": "610c5b519420a3633ec4950c15a26d83ad0e9be35ef9d00a9e3d2b621b325f72",
       "url": "https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/config.json"
     },
     "model/generation_config.json": {
       "bytes": 182,
       "sha256": "83595cf3b045d7ed16593ae399b0e3c13fe45640a4bec3a1cd371b057a6a98b7",
       "url": "https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/generation_config.json"
     },
     "model/configuration_minicpm.py": {
       "bytes": 9660,
       "sha256": "1e289912fdc0608ccad3983f4ff8b381ec8f161f45c85be4d694f92f17596b0a",
       "url": "https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/configuration_minicpm.py"
     },
     "model/modeling_minicpm.py": {
       "bytes": 69994,
       "sha256": "c646022c44d0e1cc1d149fb80422a5d5f1ba6de66dd0f525f2be1e18d99994bf",
       "url": "https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/modeling_minicpm.py"
     },
     "model/tokenizer.json": {
       "bytes": 6701371,
       "sha256": "adf7208af154a5ca065d2eda4e5419e02aac58c2c00627874748b75ec6769094",
       "url": "https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/tokenizer.json"
     },
     "model/tokenizer_config.json": {
       "bytes": 2835,
       "sha256": "8bef75d59004eb0e62db89e97090d4736b14fd6c4a0282370c28833be1e83051",
       "url": "https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/tokenizer_config.json"
     },
     "model/special_tokens_map.json": {
       "bytes": 630,
       "sha256": "9ec9d6f07cb0e436598bf29b4d97c6355434c08fdbe9c264e6926907f3751c7e",
       "url": "https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/special_tokens_map.json"
     },
     "model/added_tokens.json": {
       "bytes": 216,
       "sha256": "4760fcbf90bc193f33827ffe02f2e7ba1af1ec43644cc02ac22fdd611f6cca15",
       "url": "https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/added_tokens.json"
     },
     "model/model.safetensors": {
       "bytes": 867772648,
       "sha256": "ef74f4de1d1c580dd64e425d11f29881b9fd39f9be16a485150f727d4583181c",
       "url": "https://huggingface.co/openbmb/MiniCPM4-0.5B/resolve/5253c7fcc5e29e1cf3eacb59a58adf1ba4df8630/model.safetensors"
     }
   }
   JSONMANIFEST
   .venv/bin/python - <<'PYHASH'
   import pathlib,json,hashlib
   for name,info in json.load(open('asset-manifest.json')).items():
    p=pathlib.Path(name);digest=hashlib.sha256()
    with p.open('rb') as stream:
     for chunk in iter(lambda:stream.read(1024*1024),b''):digest.update(chunk)
    assert p.stat().st_size==info['bytes'] and digest.hexdigest()==info['sha256'],name
   print('All nine native asset checks passed.')
   PYHASH

   Expected result: All nine exact asset checks pass. Stop on an access error, byte-size mismatch or checksum mismatch.

Configure

1. 1. Create a private key

   Keep the key in this private directory. Never print it or share it. Requests and health checks require the same key.

   CommandCopy command

   set -eu
   cd "$HOME/minicpm4-05b"
   umask 077
   .venv/bin/python - <<'PYKEY'
   import pathlib,secrets
   p=pathlib.Path('.api-key')
   with p.open('x') as f:f.write(secrets.token_hex(32)+'\n')
   PYKEY

   Expected result: Private key file created with restrictive permissions; no key printed.

2. 2. Write the checked native CPU loader

   The loader streams all asset checksums, rejects all native loading diagnostics and uses full FP32 eager CPU execution. Trust is limited to the exact reviewed local MiniCPM source. Cache is explicitly disabled at generation.

   CommandCopy command

   cd "$HOME/minicpm4-05b"
   cat > reader-chat-server.py <<'PYSERVER1'
   import pathlib,json,hashlib,hmac
   from http.server import BaseHTTPRequestHandler,HTTPServer
   ROOT=pathlib.Path(__file__).resolve().parent
   PORT=8994
   
   def valid_prompt(prompt):
       return type(prompt) is str and bool(prompt.strip()) and len(prompt.encode('utf-8'))<=4096 and not any(x in prompt for x in ['<|','<s>','</s>'])
   
   def validate_completion(result):
       n,ids,answer=result
       if type(n) is not int or not 1<=n<=3584:raise ValueError('prompt_budget')
       if not isinstance(ids,list) or not 2<=len(ids)<512 or any(type(x) is not int or x<0 or x>=73448 for x in ids):raise ValueError('output_budget_or_ids')
       if ids[-1] not in (2,73440) or any(x in (2,73440) for x in ids[:-1]):raise ValueError('native_eos_required')
       if type(answer) is not str or not answer.strip() or len(answer.encode('utf-8'))>16384:raise ValueError('empty_or_excessive_answer')
       return {'answer':answer,'prompt_tokens':n,'generated_tokens_including_eos':len(ids),'complete_by_native_eos':True,'quality_measured':False}
   
   def validate_loading_info(info):
       fields={'missing_keys','unexpected_keys','mismatched_keys','error_msgs'}
       if type(info) is not dict or set(info)!=fields or any(type(info[k]) is not list or info[k] for k in fields):raise RuntimeError('incomplete_native_weight_load')
   
   def make_encoder():
       import os
       os.environ['HF_HUB_OFFLINE']='1';os.environ['TRANSFORMERS_OFFLINE']='1'
       import torch
       from transformers import AutoModelForCausalLM,PreTrainedTokenizerFast
       torch.set_num_threads(4)
       manifest=json.loads((ROOT/'asset-manifest.json').read_text())
       for name,info in manifest.items():
           file=ROOT/name;digest=hashlib.sha256()
           with file.open('rb') as stream:
               for chunk in iter(lambda:stream.read(1024*1024),b''):digest.update(chunk)
           if file.stat().st_size!=info['bytes'] or digest.hexdigest()!=info['sha256']:raise RuntimeError('asset_mismatch:'+name)
       tokenizer=PreTrainedTokenizerFast.from_pretrained(ROOT/'model',local_files_only=True)
       if tokenizer.bos_token_id!=1 or tokenizer.eos_token_id!=73440 or len(tokenizer)!=73448:raise RuntimeError('native_tokenizer_configuration')
       model,info=AutoModelForCausalLM.from_pretrained(ROOT/'model',local_files_only=True,trust_remote_code=True,
           dtype=torch.float32,attn_implementation='eager',use_safetensors=True,output_loading_info=True,low_cpu_mem_usage=True)
       validate_loading_info(info)
       cfg=model.config
       if not (cfg.model_type=='minicpm' and cfg.hidden_size==1024 and cfg.num_hidden_layers==24 and cfg.num_attention_heads==16 and cfg.num_key_value_heads==2 and cfg.max_position_embeddings==32768):raise RuntimeError('native_architecture')
       model=model.to('cpu').float().eval()
       for param in model.parameters():param.requires_grad_(False)
       template=json.loads((ROOT/'model/tokenizer_config.json').read_text())['chat_template']
       def encode(prompt):
           text=tokenizer.apply_chat_template([{'role':'user','content':prompt}],chat_template=template,tokenize=False,add_generation_prompt=True)
           inputs=tokenizer(text,add_special_tokens=True,return_tensors='pt',truncation=False)
           n=inputs['input_ids'].shape[1]
           if n<1 or n>3584 or inputs['input_ids'][0,0].item()!=1 or (inputs['input_ids']==1).sum().item()!=1:raise ValueError('prompt_budget_or_bos')
           with torch.inference_mode():
               full=model.generate(**inputs,max_new_tokens=512,do_sample=True,temperature=0.7,top_p=0.7,top_k=0,
                   repetition_penalty=1.0,num_beams=1,eos_token_id=[2,73440],pad_token_id=2,use_cache=False)
           if full.shape[0]!=1 or not torch.equal(full[0,:n],inputs['input_ids'][0]):raise ValueError('prefix_changed')
           ids=full[0,n:].tolist()
           return n,ids,tokenizer.decode(ids,skip_special_tokens=True)
       return encode
   
   PYSERVER1

   Expected result: Native loader and protocol validation written; no model loaded yet.

3. 3. Write the private serial HTTP boundary

   Complete the server. It accepts one plain prompt with strict framing, authentication and byte/token limits; template control strings are rejected. It performs no tool action, URL fetch or file upload.

   CommandCopy command

   cd "$HOME/minicpm4-05b"
   cat >> reader-chat-server.py <<'PYSERVER2'
   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, 'context_tokens': 4096,
                                       'output_cap': 512})
   
           def do_POST(self):
               if not self.authorized():
                   return self.reply(401, {'error': 'unauthorized'})
               if self.path != '/v1/chat':
                   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 > 8192:
                   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 set(request)!={'prompt'} or not valid_prompt(request['prompt']):
                   return self.reply(400, {'error': 'bounded_plain_prompt_required'})
               try:
                   result = validate_completion(encoder(request['prompt']))
               except Exception:
                   return self.reply(500, {'error': 'generation_or_protocol_failure'})
               return self.reply(200, result)
   
       return HTTPServer(('127.0.0.1', port), Handler)
   
   
   if __name__=='__main__':
       encoder=make_encoder()
       with create_server(encoder,(ROOT/'.api-key').read_text().strip()) as server:server.serve_forever()
   PYSERVER2

   Expected result: Complete private serial server written; model work starts only in the next step.

Start the model

1. 1. Start the native model process

   Keep this foreground terminal open. Startup first validates all files and native loading diagnostics. Readiness occurs only after the actual model finishes loading; stop on any failure rather than relaxing checks.

   CommandCopy command

   cd "$HOME/minicpm4-05b"
   .venv/bin/python reader-chat-server.py

   Expected result: Foreground process remains running after a successful native load. Memory fit and startup time must be observed by the reader; no result was measured here.

Verify the setup

1. 1. Check readiness, actual text completion and authentication

   Use a second terminal on the same host. Commands use the virtual environment by absolute relative path and require no prior activation. Inspect the actual answer and record semantic errors. EOS, shape and token-budget validity are separate from quality. A client timeout does not cancel native inference.

   CommandCopy command

   set -eu
   cd "$HOME/minicpm4-05b"
   MINICPM_KEY=$(cat .api-key)
   export MINICPM_KEY
   curl -fsS --max-time 30 -H "Authorization: Bearer $MINICPM_KEY" http://127.0.0.1:8994/health -o health.json
   curl -fsS --max-time 1800 -H "Authorization: Bearer $MINICPM_KEY" -H 'Content-Type: application/json' --data '{"prompt":"Describe a quiet garden in two short sentences."}' http://127.0.0.1:8994/v1/chat -o answer.json
   .venv/bin/python - <<'PYCHECK'
   import json
   h=json.load(open('health.json'));r=json.load(open('answer.json'))
   assert h=={'ready':True,'context_tokens':4096,'output_cap':512}
   assert r['complete_by_native_eos'] is True and r['quality_measured'] is False
   assert 1<=r['prompt_tokens']<=3584 and 2<=r['generated_tokens_including_eos']<512
   assert isinstance(r['answer'],str) and r['answer'].strip()
   print(r['answer'])
   print('Protocol complete. Record whether this actual answer describes a garden and follows the requested length; this is human assessment, not a forced model result.')
   PYCHECK
   code=$(curl -sS --max-time 30 -o unauthorized.json -w '%{http_code}' http://127.0.0.1:8994/health)
   test "$code" = 401
   unset MINICPM_KEY

   Expected result: Authenticated health succeeds, an actual nonempty answer terminates at a native EOS below the cap, and unauthenticated health returns 401. No exact wording or quality is promised.

Performance evidence

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

Limits and troubleshooting

• Only this exact native MiniCPM4-0.5B text checkpoint and bounded CPU path are source-qualified. Other MiniCPM4/4.1 sizes, sparse/hybrid reasoning, quantization, GPU, long-context, tool and newer-family workflows remain open.
• The native configuration has 32,768 positions; this author selects a 4,096-token budget with 512 reserved output tokens. Tokenizer metadata contains a sentinel maximum rather than a usable model limit; it is not used as context evidence. Actual prompt/output counts remain unmeasured until execution.
• The native template has no BOS; its pinned tokenizer postprocessor adds one BOS during the single tokenization pass. The wrapper rejects duplicate BOS, template-control input, truncated prompts, partial cap-length completions and every native loading diagnostic.
• Custom native Python executes only after reader review of the exact pinned files. Source compatibility is not an executed installation or load. All model work uses CPU FP32 eager operations with cache disabled, recomputing the sequence rather than making a cache-performance claim.
• The serial HTTP boundary has authentication, loopback, body and socket limits, but no generation deadline, public TLS or multiuser resilience. Socket/client timeouts do not cancel generation. Actual answer quality, fit and speed are unknown.

Asset, custom code or load diagnostic mismatch

Stop. Restore the exact pinned download and native runtime. Do not ignore missing, unexpected or mismatched keys, substitute a different model, enable network code fallback or relax checks.

Empty answer, no EOS, output cap or prompt overflow

Record the actual failure. Do not silently truncate inputs or accept partial answers. Keep the selected total budget and requalify any changed configuration.

Incorrect or unhelpful answer

Record the actual answer and human assessment. Protocol validity does not establish task quality or a benchmark score.

OOM, slow CPU or dependency error

Stop and record the actual environment and error. Planning capacity is not proven fit. Eager cache-disabled inference may be slow; any changed runtime or precision requires requalification.

401 or connection refused

Check foreground readiness, private key and loopback or SSH boundary. Keep the service port closed publicly and never print the key.

Stop the model

1. 1. Stop the private model process

   Press Ctrl+C in the foreground terminal and unset the key in each client terminal. Close the SSH tunnel; stopping Python does not stop AWS billing.

   CommandCopy command

   unset MINICPM_KEY

   Expected result: Private Python service stops and the model process exits.

Uninstall

1. 1. Remove the dedicated workspace

   Back up wanted results and license records, verify the exact directory and remove only this workspace. Shared Ubuntu packages remain installed.

   CommandCopy command

   rm -r -- "$HOME/minicpm4-05b"

   Expected result: Dedicated model, custom code, tokenizer, virtual environment, reader wrapper, outputs and private key removed.

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.