Alpha
Contact
Assess · Tucano 2

Can you own it?

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

This page is a projection of the one entry record, the Use & modify and Transparency factors that Assess 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

Intended & out-of-scope use

Adopt the Instruct and Think models as Portuguese assistants only after your own safety evaluation, and the Base models as a starting point for fine-tuning. The cards give the primary use as research and development on Portuguese language modelling, with fine-tuning for deployment allowed under Apache-2.0 (ev-card-instruct, ev-card-base). The Base cards say the models are not for deployment or human-facing use, are not fine-tuned for downstream tasks and are for Portuguese only. The Think card describes a reasoning model with no coding training (ev-card-think).

Known limitations, bias & failure modes

The cards list hallucination, social bias and toxicity from web data, weaker output in other languages, and repetition loops when the repetition penalty is low (ev-card-instruct).

The paper adds weaker instruction following and coding than Qwen3-4B. IFEval-PT is 41.67 against 79.33, and HumanEval is 47.56 against 86.59 for the 3.7B Instruct. It attributes the gap to the SFT data mix. Long-context retrieval is weak: RULER-PT at 4,096 tokens is 0.686 for the 3.7B Instruct against 0.979 for Qwen3-4B (ev-paper). A third-party run saw an invented fact about a Brazilian city and a wrong answer to who the model is (ev-community).

Openness tier & components

The tier is open_weights_recipe (ceiling 4 on the openness dimension).

ComponentGradeEvidence
Weightsopensafetensors for nine models, ungated, with Base intermediate checkpoints as branches (ev-hf-platform, ev-card-base)
Training datapartialGigaVerbo-v2 (320B tokens), Synth (9.3B), SFT (4.1M examples) and Preferences (28,437 pairs) are downloadable; the Qwen3 base data is not published (ev-gv2, ev-gv2-post, ev-qwen3-licence)
Training codeopenllm-foundry, Apache-2.0, plus YAML configs and logs in each repo (ev-code, ev-configs)
Evaluationopena Portuguese lm-evaluation-harness fork (MIT) and published result files (ev-code, ev-card-instruct)
Documentationopenmodel cards with limitations and a full paper (ev-card-instruct, ev-paper)
LicenceopenApache-2.0 (ev-licence-files)

The cards say that all data, code and recipes are open and reproducible. That holds for the Tucano stages. It does not extend to the Qwen3 stage underneath, and the pretraining corpus carries a mixed licence.

Licence terms & permitted use

Apache-2.0 governs every Tucano 2 weights repo, ungated, with commercial use allowed (ev-licence-files). Eight of the nine Qwen-derived repos carry an identical LICENSE file. Tucano2-qwen-1.5B-Base has none and declares Apache-2.0 only in its card, so add the licence text yourself when you redistribute it. The file names four individuals as copyright holders.

The upstream models are Apache-2.0: the Qwen3 bases, and the Qwen2.5 7B, 14B and 32B models that wrote the synthetic data (ev-qwen3-licence, ev-qwen25-licence). An abliterated Qwen2.5-32B variant wrote the rejected harmful responses in the preference data and is also Apache-2.0.

The pretraining corpus is the open question. Its card says the combined dataset has no single unified licence and lists sources under ODC-By, CC0, CC-BY, CC-BY-SA and two non-commercial licences, CC-BY-NC-SA-4.0 (xlsum, 14 million tokens) and CC-BY-NC-4.0 (Bactrian-X, 8 million tokens) (ev-gv2). Neither the card nor the paper says whether those texts entered the sampled training mixture, and no document states what the NC terms mean for the weights.

Supply-chain & provenance

The canonical source is huggingface.co/Polygl0t, which Hugging Face reports as not verified, with seven members (ev-hf-platform). Weights are safetensors only, each file has a Hugging Face LFS sha256, and the publisher ships no checksum file, signature or attestation. The API gives no malware-scan result.

Nine models derive from Qwen3 base models: Qwen3-0.6B-Base, 1.7B-Base and 4B-Base, with the Qwen tokenizer replaced by a 49K-token Portuguese tokenizer through Orthogonal Matching Pursuit transplantation (ev-card-base, ev-paper). They continue on 50B tokens (0.5B and 3.7B) or 100B tokens (1.5B), about 0.3% of the Qwen3 pretraining volume per the paper. The post-training data was generated by Qwen2.5 models. None of the nine is trained from scratch.

The only from-scratch release is Tucano2-0.6B-Base, a Llama-architecture 670M model trained on about 408B tokens, with no instruct variant (ev-card-06b). It is a separate follow-up entry.

EU AI Act posture

The models are general-purpose AI. The Qwen3-4B base is about 8.6e23 FLOPs by OneHill's estimate from the Qwen3 card (6 x 4B parameters x 36T tokens), and Tucano adds 0.13% of that per the paper. Both are far under the 1e25 systemic-risk threshold, so no Article 55 duties apply (ev-paper, ev-qwen3-licence).

Apache-2.0 would normally qualify for the Article 53 open-source exemption, and no monetisation is documented. The exemption never covers a copyright policy or a public training-content summary. Neither exists as a standalone document. The paper and dataset cards list sources and licences for the Tucano stages, and the Qwen3 data is described only at the level of 36T tokens in 119 languages. The paper does not mention the Act, and no code-of-practice status appears in any document. This is guidance, not legal advice.

Benchmarks & evaluation

All figures are the publisher's own, on a harness it released. Total scores are Normalised Performance Metric averages over the publisher's Portuguese suite (ev-card-base, ev-card-instruct, ev-card-think, ev-paper).

ModelResultReference points in the cards
3.7B Base, total59.21Qwen2.5-7B 57.97, Qwen3-4B-Base 57.86
3.7B Instruct, total53.64 (Knowledge and Reasoning 56.22, IFEval-PT 41.67, HumanEval 47.56)Jurema-7B 53.03, Qwen2.5-3B-Instruct 51.71, Qwen3-4B 51.36
3.7B Instruct, single tasksENEM 72.92, BLUEX 64.53, OAB 54.31, Belebele 85.22, MMLU 64.64, GSM8K-PT 53.81Qwen3-4B MMLU 26.93
3.7B Think, total51.27 (Knowledge and Reasoning 54.07)SmolLM3-3B 48.58, Qwen3-4B 46.35
1.5B Instruct, total37.54Qwen3-1.7B 36.30
0.5B Instruct, total26.08Qwen3-0.6B 22.21

One third-party run on 393 text-only ENEM questions gave 74.6% (95% confidence interval 70.0 to 78.6) for the 3.7B Instruct, ahead of Qwen3-8B in W4A16 at 67.9% (ev-community). The author warns that ENEM may be in training data. No independent study of the full suite exists in the sources found. The Qwen3-4B MMLU figure of 26.93 in the Instruct card is far below the 3B and 1.7B rows, which suggests a harness artefact. That reading is OneHill's inference, so treat comparisons on that task with care.

Score summary

Rubric 1.1, computed by scripts/score.py from the dimension scores in entry.yaml: openness 4, provenance 2, legal 3, safety 2, performance 2, operational 3, governance 3. The reasons are in each dimension rationale in entry.yaml.

Ownership verdict

Level: partial. Use and modify is strong, data control is strong, transparency is moderate, and reliability is weak because the safety dimension is rated 2. The ownership block in entry.yaml gives the verdict text and the conditions that would raise it.

Contradictions and surprises

The paper and the Base cards call the whole suite fully open, reproducible and released under permissive licences. The GigaVerbo-v2 card says the corpus has no unified licence and lists two non-commercial sources, and the Qwen3 base data is not published. The accurate term for the suite is open weights with an open recipe.

The paper abstract gives sizes of 0.5 to 3.7B and the introduction says 0.5B to 3.5B. The cards report 490,799,104, 1,510,073,344 and 3,759,341,056 parameters.

Tucano2-qwen-1.5B-Base has no LICENSE file while its siblings have one. The Hugging Face organisation is unverified although the paper names its authors and their University of Bonn affiliations.

How this scores

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

1

Use and modify freelyCan you run, modify and adapt it with no gate and no field-of-use trap?

Strong

Apache-2.0 LICENSE files, ungated weights and open training code, configs and data let you run, modify and redistribute the models commercially (ev-licence-files, ev-configs, ev-code). Tucano2-qwen-1.5B-Base lacks a LICENSE file.

How this scores (AOI sub-dimensions)
Openness4/5how much is released - weights, data, code, licence - and how freelyWeights, training code (llm-foundry), configs, logs, evaluation harness and four datasets (GigaVerbo-v2, Synth, SFT, Preferences) are public and ungated under Apache-2.0 or other stated licences.
Legal3/5how permissive and clean the licence is for real commercial useApache-2.0 LICENSE files, a stated commercial-use grant, and a paper that lists sources, generators, compute and energy support common uses.
2

TransparencyDo you know what it is: weights, training, behaviour, and legible terms?

Moderate

The Tucano stages are documented down to configs, logs and datasets (ev-paper, ev-configs). The Qwen3 base stage and its 36T-token data are undisclosed (ev-qwen3-licence), and the corpus carries no single licence (ev-gv2).

How this scores (AOI sub-dimensions)
Provenance2/5how well we can trace and verify what went into the modelSafetensors only, with a Hugging Face LFS sha256 per file, and a clear lineage to Qwen3 base models.
Governance3/5how accountable and well-documented the publisher isA named University of Bonn team with a funded project (BMBF and the NRW science ministry), an arXiv paper, a prior Tucano series and a 2026-08-13 update to the post-trained repos.
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.