Alpha
Contact
Implement · Tucano 2

Can you run it?

Ownership levelPartialnone·limited·partial·substantial·fullAnalytical input D · 54.4/100

This page is a projection of the one entry record, the Reliability and Data control factors that Implement covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.

Which domain expands which factor
  • AssessUse & modify + Transparency
  • ImplementData control + Reliability
  • UseReliability
  • SupportTransparency

Install & run

The models load with the standard transformers classes, with no remote code. Pin a revision hash and record the LFS sha256 of each safetensors shard (ev-card-instruct, ev-configs).

from transformers import AutoTokenizer, AutoModelForCausalLM
model_id = "Polygl0t/Tucano2-qwen-3.7B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
messages = [{"role": "user", "content": "Qual é a capital de Cabo Verde?"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

The cards set temperature 0.1, top_k 50, top_p 1.0 and repetition_penalty 1.2, and generation_config.json sets max_new_tokens to 1024.

Hardware & VRAM requirements

The publisher gives file sizes and training cost, not a requirements table (ev-hf-platform).

Modelbf16 weights
3.7B Base, Instruct, Think7.77 GB (Base), 7.52 GB (Instruct, Think) in two shards
1.5B3.02 GB
0.5B0.98 GB

A third party served the 3.7B Instruct in bf16 with vLLM on an RTX 3080 (about 9.6 GiB usable). The setup used a 4,096 context, a 1 GiB KV cache and 8.7 GiB of VRAM. Speed was 79.4 tokens per second for one request and 436.7 at concurrency 8 (ev-community). OneHill has not measured memory or speed.

Serving stacks

StackStatus
transformersdocumented on every card (ev-card-instruct)
vLLMrun by a third party with the unmodified Instruct weights (ev-community)
MLXcommunity 4-bit build of the 3.7B Instruct (ev-community)
llama.cpp, Ollama, TGIno documented support or GGUF; the Qwen3 architecture is standard, but OneHill has not tested it
Hosted providersnone listed by Hugging Face (ev-hf-platform)

Safe-deployment controls & Deployment Ceiling

Apply the safe-deployment playbook.

  • Supply chain: pin the revision and record the LFS hashes, since no checksum file or signature ships.
  • Model layer: use the Instruct or Think models only. The Base models have no safety tuning and the cards bar deployment of them.
  • Evaluation: run your own Portuguese safety, bias and toxicity tests. No safety evaluation is published.
  • Input and output layers: add guardrails. No guard model or acceptable-use policy ships.
  • Context: keep prompts and context under 4,096 tokens.
  • Licence hygiene: add the Apache-2.0 text to anything you redistribute, and record the mixed GigaVerbo-v2 source licences in your AI-BOM.

Deployment Ceiling: T1 (internal assistant, a human reads every output), conditional. The evidence read does not support customer-facing T2 use because no safety evaluation exists.

Available quantizations

The publisher ships none. A community MLX 4-bit build of the 3.7B Instruct exists (pessini/Tucano2-qwen-3.7B-Instruct-MLX-4bit, Apache-2.0). Quality loss in Portuguese is unmeasured, so test your own tasks before using it.

Fine-tuning & adaptation

The repos publish the SFT and APO configs, training logs and the llm-foundry trainer, so you can rerun or adapt the post-training stages (ev-configs, ev-code). The Instruct SFT run uses 5 epochs at a peak learning rate of 5e-5 with assistant-only loss and a batch of 524,288 tokens. The APO run uses 5 epochs at 5e-6, apo_zero loss and beta 0.5, on 28,437 preference pairs. The Base repos keep intermediate checkpoints as branches every 2,500 steps for the 3.7B model. The SFT and Preferences datasets are Apache-2.0 and ungated. Check the CC-BY-NC question in the corpus licence before using GigaVerbo-v2 for commercial continued pretraining.

API / OpenAI-compatible integration

There is no hosted endpoint. A third-party recipe serves the 3.7B Instruct through vLLM's OpenAI-compatible chat completions route (ev-community). The chat template carries a tools block in Portuguese, so pass tools through the template and not as free text.

How this scores

The ownership factors this domain covers, drawn from the one entry record.

3

ReliabilityIs it reliable and good enough for the job?

Weak

Safety is rated 2 because no safety evaluation, acceptable-use policy or guard model is published (ev-card-instruct, ev-paper). Operational readiness is 3, with no official quantisation or serving guide.

How this scores (AOI sub-dimensions)
Operational3/5how practical it is to run, serve and maintain in productionThe weights use the standard Qwen3 architecture in transformers with safetensors.
Safety2/5whether misuse risks are evaluated and guardrails are providedThe Instruct and Think models are safety-tuned with a published 28,437-pair preference set built from written constitutions, and the cards list hallucination, bias, toxicity and repetition.
4

Doesn't extract your dataDoes running it keep your knowledge and data yours?

Strong

The weights run on your own hardware under Apache-2.0 with no clawback clause, and the architecture is the standard Qwen3 class with no remote code (ev-licence-files, ev-configs).

How this scores
Not a scored AOI dimension. For a self-hosted model, data-control is a structural property of running the weights yourself, strong by default unless the model phones home or the licence claws back rights. For a hosted API this factor is the retention + train-on-inputs + residency read, scored from the binding terms.
What this means for adoptionYou own the weights: Apache-2.0, ungated, commercial use allowed, with the code and data to retrain the Tucano stages on your own hardware. You do not own a verified safe assistant, because no safety evaluation exists, and you cannot rebuild the Qwen3 stage underneath. The level rises if Polygl0t publishes a safety evaluation and resolves the mixed licence of the training corpus.

Sources

The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.

Model cardread2026-11-19
The 3.7B Instruct card declares Apache-2.0, 3,759,341,056 parameters and a 4,096-token context, and describes one SFT round and one APO round on gigaverbo-v2-sft and gigaverbo-v2-preferences.
Model cardread2026-11-19
The 3.7B Base card says the model is continually pretrained from Qwen3-4B-Base on about 50 billion tokens, with the Tucano2 tokenizer transplanted by Orthogonal Matching Pursuit.
Model cardread2026-11-19
The 3.7B Think card declares Apache-2.0 and describes a reasoning model trained with SFT and APO that writes chain-of-thought inside think tags.
Model cardread2026-11-19
The Tucano2-0.6B-Base card declares Apache-2.0, a Llama architecture with 670,127,616 parameters, about 400 billion training tokens and three published stage configs.
Licenceread2026-11-19
Eight of the nine Qwen-derived repos and the 0.6B Base repo carry an identical Apache-2.0 LICENSE file that names four individuals as copyright holders.
Licenceread2026-11-19
Qwen3-4B-Base, 1.7B-Base and 0.6B-Base each carry an Apache-2.0 LICENSE file and are ungated.
Licenceread2026-11-19
Qwen2.5-7B, 14B and 32B-Instruct, which generated the synthetic data, and the abliterated 32B variant that wrote the rejected responses, are Apache-2.0 with LICENSE files and ungated.
Technical_reportread2026-11-19
The paper 'Tucano 2 Cool' (arXiv 2603.03543, 3 March 2026) describes continual pretraining from Qwen3 bases, a from-scratch 0.6B model, the data, the post-training recipe and the publisher's benchmarks.
Documentationread2026-11-19
The GigaVerbo-v2 card lists its source datasets with their licences and states that the combined dataset has no single unified licence (license: other).
Documentationread2026-11-19
The Synth, SFT and Preferences cards declare Apache-2.0 and are ungated.
Documentationread2026-11-19
llm-foundry (training, data, alignment and evaluation code) is Apache-2.0 and the Portuguese lm-evaluation-harness fork is MIT.
Documentationread2026-11-19
The Polygl0t Hugging Face organisation is not verified and has seven members.
Documentationread2026-11-19
The Instruct and Think repos publish SFT and APO YAML configs, training logs, config.json and a ChatML chat template.
Third-party analysisread2026-11-19
A third-party repo ran the 3.7B Instruct in bf16 on vLLM 0.33.0 on an RTX 3080 with 8.7 GiB of VRAM and scored 74.6% on 393 text-only ENEM questions.