Local hardwareDocumented setupSources checked Oct 16, 2026
Run Manacá raw text continuation on a 32 GiB Intel desktop
Serve the exact 1.72B-parameter Brazilian-Portuguese base model using native Transformers, its corrected lowercase Unigram tokenizer and CPU FP32. A reader-authored private completion endpoint accepts raw text and returns only newly generated text. Use four threads, one request, 4,096 total tokens and a 40-token greedy check. 32 GiB RAM and 40 GiB disk are planning allocations; full fit and performance are unknown.
- Hardware
- Intel Core i5-12400 desktop, 4 inference threads, 32 GiB RAM
- Runtime
- Native Transformers Llama with reader-authored private raw-completion service Transformers 4.57.6; torch 2.6.0+cpu cp312; Python 3.12; author-selected FP32 · CPU
- Model format
- safetensors · CPU FP32 runtime from native BF16 checkpoint · No quantization; native BF16 checkpoint converted to FP32 in CPU memory
- Workload
- 4,096 token context · 1 concurrent
Based on source documentation; this setup has not been executed by OpenTools. Memory: unknown. No compatible measurement available.
Read the complete setup guide
Before you start
- Use an owned Intel Core i5-12400 desktop with Ubuntu 24.04 x86_64, 32 GiB installed RAM and 40 GiB free disk. Allocate four inference threads and build natively on the same host.
- Use exact native model/tokenizer files, Python 3.12, Transformers 4.57.6 and the hashed torch 2.6.0 CPU wheel. No remote model code or GPU driver is used.
- The publisher model is base pretrained Brazilian-Portuguese text continuation, not instruction-tuned chat. Use the corrected native Unigram tokenizer, which normalizes NFKC then lowercase.
- Read the pinned CC BY 4.0 declarations and official project LICENSE; retain attribution, license link and change notices. 32 GiB RAM and 40 GiB disk are allocations, not measured fit.
Exact model and requirements
- Artifact:
- AIInstituteLNCC (catalog repository redirects from menezesbruno) / Manacá-1B Base native safetensors (1,722,951,680 publisher parameters) · checkpoint AIInstituteLNCC/manaca-1b-base
- Publisher revision:
- a9f5da7cb220ec4b0205a45da2455946c44e812f
- Artifact digest:
- sha256:9edc4dc112199a8c7dd73b0e10a70d69cf8e46e42e4db5505051935451fa6648
- Download size:
- 3.21 GiB
- Memory requirement (unknown):
- Unknown — No full peak measured for exact native Manacá CPU FP32 model, tokenizer, cache and wrapper.
- System RAM / disk:
- Unknown / Unknown
- Placement:
- single device. One CPU host; no tensor/pipeline sharding. Entire model on CPU; no device_map, GPU or disk offload.
- KV cache:
- FP32 native attention K/V under the explicitly FP32 CPU model; full peak memory unmeasured.
- Workload:
- 4,096 token context; Unknown; Unknown; concurrency 1; batch 1; cache unknown; repetitions Unknown.
- Tokenizer:
- Exact native fast Unigram tokenizer with NFKC then Lowercase and native BOS · a9f5da7cb220ec4b0205a45da2455946c44e812f
- Sampling / reasoning:
- Greedy do_sample=False and max_new_tokens 40 follow the publisher raw-continuation example. Wrapper permits 1–64 new tokens and enforces input plus output <=4,096; repetition_penalty 1 is an explicit author choice. Pretrained base raw continuation; no chat template, instruction role formatting, thinking controls or tool executor. Newly generated token IDs are decoded separately from input.
- Runtime prerequisites:
- Ubuntu 24.04 x86_64; See prerequisites above.
Exact model README declares CC BY 4.0 and links the official project, whose pinned LICENSE states attribution, license-link and change-notice obligations and no additional restrictions. Root model LICENSE returned 404 and Creative Commons legalcode requests returned 403; these failures are preserved. No OSI approval or legal-clearance claim. Retain appropriate creator credit and applicable notices when sharing.
- config.json · 0 GiB · revision a9f5da7cb220ec4b0205a45da2455946c44e812f
- generation_config.json · 0 GiB · revision a9f5da7cb220ec4b0205a45da2455946c44e812f
- tokenizer.json · 0 GiB · revision a9f5da7cb220ec4b0205a45da2455946c44e812f
- tokenizer_config.json · 0 GiB · revision a9f5da7cb220ec4b0205a45da2455946c44e812f
- special_tokens_map.json · 0 GiB · revision a9f5da7cb220ec4b0205a45da2455946c44e812f
- The original native artifact is 3,445,946,840 bytes. Converting BF16 weights to FP32 increases weight memory; attention cache, buffers, OS and dependencies add memory. File size is not full fit.
- 32 GiB RAM, 40 GiB disk and four threads are author planning allocations, not measured minimums. No sibling model fit or performance is transferred.
One-time setup and hardware
- hardware
- Unknown — Existing desktop purchase price and incremental hardware cost depend on the reader; unknown.
- Existing desktop purchase price and incremental hardware cost depend on the reader; unknown. Tax: unknown.
- setup
- Unknown — Initial provisioning/build time and labor cost are unknown.
- Initial provisioning/build time and labor cost are unknown. Tax: unknown.
Recurring costs
- electricity
- Unknown — Actual wall power, electricity rate and powered hours are unknown; no TDP-to-wall-power conversion.
- Actual wall power, electricity rate and powered hours are unknown; no TDP-to-wall-power conversion. Tax: unknown.
- storage
- Unknown — Incremental cost of local disk capacity is unknown. The 40 GiB free-disk walkthrough allocation is an author choice, not a measured requirement.
- Incremental cost of local disk capacity is unknown. The 40 GiB free-disk walkthrough allocation is an author choice, not a measured requirement. Tax: unknown.
- egress
- Unknown — Data transfer price, allowances and traffic volume are unknown.
- Data transfer price, allowances and traffic volume are unknown. Tax: unknown.
- administration
- Unknown — Maintenance labor rate and hours are unknown.
- Maintenance labor rate and hours are unknown. Tax: unknown.
Install the runtime
1. Create the Python 3.12 CPU environment
Install the exact official CPU wheel by hash and Transformers version. System and transitive package versions resolve at installation; save the resulting environment. This is not an executed compatibility claim.
Commandsudo apt-get update sudo apt-get install -y python3.12 python3.12-venv curl ca-certificates openssl mkdir -p "$HOME/manaca-hosting/model" cd "$HOME/manaca-hosting" 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" .venv/bin/python -m pip check .venv/bin/python -m pip freeze > resolved-environment.txtExpected result: pip check succeeds; resolved-environment.txt records actual installed versions. Stop on resolver or hash errors.
Download the model
1. Download exact native model and tokenizer inputs
Verify original full safetensors and every config/tokenizer auxiliary before loading. Do not replace the corrected native tokenizer. No chat-template file is supplied.
Commandcd "$HOME/manaca-hosting" curl --fail --location --output model/model.safetensors 'https://huggingface.co/AIInstituteLNCC/manaca-1b-base/resolve/a9f5da7cb220ec4b0205a45da2455946c44e812f/model.safetensors' curl --fail --location --output model/config.json 'https://huggingface.co/AIInstituteLNCC/manaca-1b-base/raw/a9f5da7cb220ec4b0205a45da2455946c44e812f/config.json' curl --fail --location --output model/generation_config.json 'https://huggingface.co/AIInstituteLNCC/manaca-1b-base/raw/a9f5da7cb220ec4b0205a45da2455946c44e812f/generation_config.json' curl --fail --location --output model/tokenizer.json 'https://huggingface.co/AIInstituteLNCC/manaca-1b-base/raw/a9f5da7cb220ec4b0205a45da2455946c44e812f/tokenizer.json' curl --fail --location --output model/tokenizer_config.json 'https://huggingface.co/AIInstituteLNCC/manaca-1b-base/raw/a9f5da7cb220ec4b0205a45da2455946c44e812f/tokenizer_config.json' curl --fail --location --output model/special_tokens_map.json 'https://huggingface.co/AIInstituteLNCC/manaca-1b-base/raw/a9f5da7cb220ec4b0205a45da2455946c44e812f/special_tokens_map.json' sha256sum --check <<'SHA256' 9edc4dc112199a8c7dd73b0e10a70d69cf8e46e42e4db5505051935451fa6648 model/model.safetensors 960455d76cfde168f01275bb7f5ecb537f556d7ab43e0968caf962e3eb2b8a54 model/config.json fab3c8a907d34bf38e431f9481082e70b0d908549f0ca6598346cf4246c4f8a3 model/generation_config.json 480e672df73cf57b4ea3ef1318480a1e0e8043bf2c2715199347eb6b57dd5efc model/tokenizer.json 3454edc64167105fbeb494d0767bf205d017a3613a83931ed480f48f883dd917 model/tokenizer_config.json 358c249e2fb29060c6b73157d428853b0c48710deffc8ee670ab1013880946c9 model/special_tokens_map.json SHA256Expected result: All six files report OK. Model is 3,445,946,840 bytes; stop on any failed download or digest mismatch.
Configure
1. Create a private service key
Create a local service credential only when absent; preserve an existing key. No provider API credential is needed. Never print or share it.
Commandcd "$HOME/manaca-hosting" umask 077 if [ ! -e .api-key ]; then openssl rand -hex 32 > .api-key; fi test -s .api-key || { echo "Restore empty service key before proceeding"; exit 1; } chmod 600 .api-keyExpected result: Nonempty private .api-key exists; previous credential preserved.
2. Create the reader-authored raw-completion service
This bounded single-request educational service is part of this walkthrough, not a publisher server. It binds loopback, authenticates health and generation, limits JSON bodies and token budget, and checks native tokenizer normalization and model bias support at startup.
Commandcd "$HOME/manaca-hosting" cat > manaca_server.py <<'PYSERVER' """OpenTools reader-authored educational demonstrator; not publisher server.""" import hmac,json,os from http.server import BaseHTTPRequestHandler,HTTPServer def check_budget(count,limit): if count>4032 or count+limit>4096: raise ValueError('token_budget') def handler_factory(generate,key): class Handler(BaseHTTPRequestHandler): protocol_version='HTTP/1.0' def setup(self): super().setup(); self.connection.settimeout(10) def log_message(self,*args): pass def reply(self,status,data): body=json.dumps(data,ensure_ascii=False).encode() self.send_response(status); self.send_header('Content-Type','application/json; charset=utf-8') self.send_header('Content-Length',str(len(body))); self.end_headers(); self.wfile.write(body) def authorized(self): value=self.headers.get('Authorization','').encode() return hmac.compare_digest(value,('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'}) self.reply(200,{'status':'ready'}) def do_POST(self): if not self.authorized(): return self.reply(401,{'error':'unauthorized'}) if self.path!='/v1/manaca': return self.reply(404,{'error':'not_found'}) if self.headers.get('Transfer-Encoding') is not None: return self.reply(400,{'error':'transfer_encoding_unsupported'}) if self.headers.get_content_type()!='application/json': return self.reply(415,{'error':'json_required'}) lengths=self.headers.get_all('Content-Length',[]) if len(lengths)!=1 or not lengths[0] or not all('0'<=c<='9' for c in lengths[0]): return self.reply(411,{'error':'content_length_required'}) if len(lengths[0])>4: return self.reply(413,{'error':'body_limit'}) n=int(lengths[0]) if not 1<=n<=8192: return self.reply(413,{'error':'body_limit'}) try: raw=self.rfile.read(n) if len(raw)!=n: return self.reply(400,{'error':'incomplete_body'}) data=json.loads(raw) except (ValueError,UnicodeError,TimeoutError): return self.reply(400,{'error':'invalid_json'}) if type(data)!=dict or set(data)!={'prompt','max_new_tokens'} or type(data['prompt'])!=str or not data['prompt'].strip() or type(data['max_new_tokens'])!=int or not 1<=data['max_new_tokens']<=64: return self.reply(400,{'error':'invalid_fields'}) try: result=generate(data['prompt'],data['max_new_tokens']) except ValueError: return self.reply(422,{'error':'token_budget'}) except Exception: return self.reply(500,{'error':'generation_failed'}) self.reply(200,result) return Handler def main(): import torch from transformers import AutoTokenizer,AutoModelForCausalLM key=open('.api-key',encoding='ascii').read().strip() if len(key)!=64 or any(c not in '0123456789abcdef' for c in key): raise SystemExit('Invalid service key') torch.set_num_threads(4) tok=AutoTokenizer.from_pretrained('./model',local_files_only=True,trust_remote_code=False,use_fast=True) assert tok.is_fast backend=json.loads(tok.backend_tokenizer.to_str()) assert backend['model']['type']=='Unigram' assert backend['normalizer']=={'type':'Sequence','normalizers':[{'type':'NFKC'},{'type':'Lowercase'}]} assert tok('A CAPITAL DO BRASIL É',add_special_tokens=False)['input_ids']==tok('a capital do brasil é',add_special_tokens=False)['input_ids'] model=AutoModelForCausalLM.from_pretrained('./model',local_files_only=True,trust_remote_code=False,dtype=torch.float32).to('cpu').eval() cfg=model.config assert cfg.model_type=='llama' and cfg.num_hidden_layers==24 and cfg.hidden_size==2048 assert cfg.attention_bias and cfg.mlp_bias and cfg.max_position_embeddings==4096 assert model.model.layers[0].self_attn.q_proj.bias is not None assert model.model.layers[0].mlp.gate_proj.bias is not None assert next(model.parameters()).dtype==torch.float32 def generate(prompt,limit): encoded=tok(prompt,return_tensors='pt',add_special_tokens=True) ids=encoded['input_ids'] count=ids.shape[-1] check_budget(count,limit) with torch.inference_mode(): out=model.generate(ids,attention_mask=torch.ones_like(ids),max_new_tokens=limit,do_sample=False,repetition_penalty=1.0,pad_token_id=tok.eos_token_id) new=out[0,count:] return {'text':tok.decode(new,skip_special_tokens=True),'input_tokens':count,'output_tokens':len(new),'eos_reached':bool(len(new) and int(new[-1])==tok.eos_token_id),'output_cap_reached':len(new)==limit} HTTPServer(('127.0.0.1',8997),handler_factory(generate,key)).serve_forever() if __name__=='__main__': main() PYSERVER .venv/bin/python -m py_compile manaca_server.pyExpected result: Python syntax check succeeds; no model is loaded by py_compile.
Start the model
1. Start one CPU raw-completion service
Run in a foreground terminal. Startup loads local files in CPU FP32, asserts fast Unigram normalization, case-equivalent tokenization, native 24-layer config and attention/MLP biases. Any failed check must stop startup. One HTTPServer processes one request at a time; do not treat this as a production server.
Commandcd "$HOME/manaca-hosting" .venv/bin/python manaca_server.pyExpected result: After loading, authenticated /health returns ready. Full fit and startup compatibility are documented, not executed here.
Connect to your endpoint
http://127.0.0.1:8997/v1/manaca
Same desktop loopback only; trusted private host.
Access: loopback. Reader-generated private Bearer key required for health and completion. Exact JSON fields prompt:string and max_new_tokens:integer 1–64. This is a reader-authored raw completion schema, not chat or a provider API.
Verify the setup
1. Check native raw continuation and access control
Use a second terminal on the server. This lowercased raw prompt follows the publisher example. Confirm output excludes the echoed prompt and returns text, input_tokens, output_tokens, eos_reached and output_cap_reached. The input count includes native BOS. A base model may reach its requested cap without EOS; that is a bounded continuation, not an instruction-answer failure or a completed-document guarantee. Inspect generated text; no exact word or factual reliability is promised.
Commandcd "$HOME/manaca-hosting" MANACA_SERVICE_KEY="$(cat .api-key)" curl --fail http://127.0.0.1:8997/health -H "Authorization: Bearer $MANACA_SERVICE_KEY" curl --fail http://127.0.0.1:8997/v1/manaca -H "Authorization: Bearer $MANACA_SERVICE_KEY" -H 'Content-Type: application/json' -d '{"prompt":"a capital do brasil é","max_new_tokens":40}' curl -s -o /dev/null -w '%{http_code}\n' http://127.0.0.1:8997/health unset MANACA_SERVICE_KEYExpected result: Authorized health ready; response has continuation-only text and numeric token counts, 1–40 output IDs, explicit EOS/cap booleans. Missing key returns 401. No model execution outcome is claimed in this guide.
Performance evidence
No compatible measurement available. Documented support and estimated memory do not establish response speed.
Limits and troubleshooting
- Documented only; no weights, installation, build or inference executed. Full memory fit, latency, speed and continuation quality remain unknown.
- Manacá has 1,722,951,680 parameters despite its rounded 1B name; native Llama attention and MLP biases must remain enabled.
- Use the exact fast tokenizer with NFKC and Lowercase. The tokenizer model_max_length sentinel is not the model limit; enforce native 4,096 combined input/output tokens. The input count includes native BOS.
- Greedy 40-token continuation follows publisher example. No instruction-following, chat-template, thinking or exact-answer guarantee. Decoding returns only new IDs; EOS and cap flags distinguish a naturally stopped versus capped continuation.
- Native weights are BF16; this CPU walkthrough explicitly converts to FP32. No quantized model or generic BF16 CPU speed claim is transferred.
- The reader-authored HTTPServer is a serial educational private endpoint: one request at a time, 8,192-byte body limit, 10-second socket-read timeout, strict JSON fields and token budget. No generation timeout, multiprocess resilience, TLS or public-service hardening is supplied.
- System and transitive dependencies resolve at installation. Record pip freeze and pip check; exact installed runtime compatibility remains unexecuted.
- 32 GiB RAM, 40 GiB disk and four threads are allocations, not measured minima. Original dated AWS compute/storage excludes tax and unknown egress/admin; local power/purchase unknown.
- Native README and linked official project LICENSE declare CC BY 4.0. Root LICENSE404 and CC legalcode403 are preserved; no legal-clearance or OSI claim. Broader workflow/backend/OS coverage remains open.
- Tokenizer/config startup assertion fails
- Stop. Recheck all exact native auxiliary hashes and native fast tokenizer normalization. Do not invent a chat template or remove bias assertions.
- HTTP401,415,411,413 or422
- Use the matching private key, JSON Content-Type and a single valid Content-Length. Keep the body <=8,192 bytes and prompt plus requested output <=4,096 native tokens; max_new_tokens must be 1–64.
- Empty or unhelpful continuation
- Inspect EOS/cap flags and decoded new token IDs. This is a base model; do not test it as an instruction chatbot. No exact completion is guaranteed.
- Out of memory or connection refused
- Inspect foreground startup and available host RAM. Stop on load failure. Resource or precision changes form a newly documented configuration.
Stop the model
1. Stop the service
Press Ctrl+C in the server terminal. Unset the key in each client shell. Confirm loopback port no longer serves.
Commandunset MANACA_SERVICE_KEYExpected result: The foreground service stops; loopback port 8997 no longer serves.
Uninstall
1. Remove dedicated files
After confirming the path and preserving wanted logs, remove only this walkthrough directory. Shared apt packages remain.
Commandrm -r -- "$HOME/manaca-hosting"Expected result: Dedicated model, environment, wrapper and private key removed.
Sources and verification
Sources checked Oct 16, 2026. Verification source
- https://huggingface.co/api/models/menezesbruno/manaca-1b-basepublisher · retrieved Oct 16, 2026
- https://huggingface.co/AIInstituteLNCC/manaca-1b-base/raw/a9f5da7cb220ec4b0205a45da2455946c44e812f/README.mdpublisher · retrieved Oct 16, 2026 · revision a9f5da7cb220ec4b0205a45da2455946c44e812f
- https://huggingface.co/AIInstituteLNCC/manaca-1b-base/raw/a9f5da7cb220ec4b0205a45da2455946c44e812f/config.jsonpublisher · retrieved Oct 16, 2026 · revision a9f5da7cb220ec4b0205a45da2455946c44e812f
- https://huggingface.co/AIInstituteLNCC/manaca-1b-base/raw/a9f5da7cb220ec4b0205a45da2455946c44e812f/tokenizer_config.jsonpublisher · retrieved Oct 16, 2026 · revision a9f5da7cb220ec4b0205a45da2455946c44e812f
- https://huggingface.co/AIInstituteLNCC/manaca-1b-base/raw/a9f5da7cb220ec4b0205a45da2455946c44e812f/generation_config.jsonpublisher · retrieved Oct 16, 2026 · revision a9f5da7cb220ec4b0205a45da2455946c44e812f
- https://huggingface.co/api/models/AIInstituteLNCC/manaca-1b-base/tree/a9f5da7cb220ec4b0205a45da2455946c44e812f?recursive=false&expand=falsepublisher · retrieved Oct 16, 2026 · revision a9f5da7cb220ec4b0205a45da2455946c44e812f
- https://huggingface.co/AIInstituteLNCC/manaca-1b-base/raw/a9f5da7cb220ec4b0205a45da2455946c44e812f/tokenizer.jsonpublisher · retrieved Oct 16, 2026 · revision a9f5da7cb220ec4b0205a45da2455946c44e812f
- https://huggingface.co/AIInstituteLNCC/manaca-1b-base/raw/a9f5da7cb220ec4b0205a45da2455946c44e812f/special_tokens_map.jsonpublisher · retrieved Oct 16, 2026 · revision a9f5da7cb220ec4b0205a45da2455946c44e812f
- https://api.github.com/repos/Instituto-IA-LNCC/manaca-1b-base/commits/mainpublisher · retrieved Oct 16, 2026
- https://raw.githubusercontent.com/huggingface/transformers/v4.57.6/src/transformers/models/llama/modeling_llama.pyruntime maintainer · retrieved Oct 16, 2026 · revision v4.57.6
- https://raw.githubusercontent.com/huggingface/transformers/v4.57.6/src/transformers/tokenization_utils_fast.pyruntime maintainer · retrieved Oct 16, 2026 · revision v4.57.6
- https://raw.githubusercontent.com/Instituto-IA-LNCC/manaca-1b-base/408cba1fb765cdc496df2a518fe32f192270e5ee/LICENSEpublisher · retrieved Oct 16, 2026
- https://raw.githubusercontent.com/Instituto-IA-LNCC/manaca-1b-base/408cba1fb765cdc496df2a518fe32f192270e5ee/README.mdpublisher · retrieved Oct 16, 2026
- https://download.pytorch.org/whl/cpu/torch/runtime maintainer · retrieved Oct 16, 2026
- https://raw.githubusercontent.com/huggingface/transformers/v4.57.6/docs/source/en/installation.mdruntime maintainer · retrieved Oct 16, 2026 · revision v4.57.6
- https://raw.githubusercontent.com/huggingface/transformers/v4.57.6/docs/source/en/models.mdruntime maintainer · retrieved Oct 16, 2026 · revision v4.57.6
- https://docs.python.org/3.11/library/http.server.htmlruntime maintainer · retrieved Oct 16, 2026
- https://docs.python.org/3.11/library/hmac.htmlruntime maintainer · retrieved Oct 16, 2026
- https://opentools.ai/llms/manaca-1b-baseopentools · retrieved Oct 16, 2026 · revision reader-raw-completion-v1
- https://www.intel.com/content/www/us/en/products/sku/134586/intel-core-i512400-processor-18m-cache-up-to-4-40-ghz/specifications.htmlmanufacturer · retrieved Oct 16, 2026
- https://raw.githubusercontent.com/ggml-org/llama.cpp/620d2225cc436ab0b001e3200e590e3b286ef9ab/ggml/src/ggml-cpu/CMakeLists.txtruntime maintainer · retrieved Oct 16, 2026 · revision 620d2225cc436ab0b001e3200e590e3b286ef9ab
- https://docs.liquid.ai/lfm/inference/llama-cppruntime maintainer · retrieved Oct 16, 2026
- https://raw.githubusercontent.com/ggml-org/llama.cpp/620d2225cc436ab0b001e3200e590e3b286ef9ab/docs/build.mdruntime maintainer · retrieved Oct 16, 2026 · revision 620d2225cc436ab0b001e3200e590e3b286ef9ab