Gervásio 8B PTPT
Gervásio 8B PTPT is a LLaMA 3.1 8B Instruct fine-tune from PORTULAN/University of Lisbon, trained on Portuguese-specific datasets including extraGLUE-Instruct and MMLU PT. It targets European Portuguese (PT-PT), not Brazilian Portuguese, so BR users should expect some dialect drift. MIT-licensed and commercially usable.
If your users write Brazilian Portuguese, this is not the right first pick — it was explicitly built for Portugal Portuguese and there is no public evidence it handles PT-BR well. The LLaMA 3.1 base is good, but the fine-tune dataset and community adoption are both small, so you are largely trusting the vendor's curation. Skip it for a BR-focused product unless you can run your own evals first. If you do need PT-PT specifically, it is worth a test given the MIT license and zero legal friction.
›Why this rating
Auto-generated rating (Opus 4.7 judge, claude-opus-4-7). Overall 9.05/10. License is explicitly MIT on the card and correctly flagged commercial-OK. Metadata (LLaMA 3.1 8B base, 32 layers, hidden 4096) matches the card; family 'llama' and parameterCountB 8 are accurate. Context length of 4096 is a reasonable inference for the row but the card excerpt doesn't explicitly confirm it — minor concern since LLaMA 3.1 base supports 128K and PORTULAN may have kept that. The useCases tag includes 'portuguese-br' which is misleading given the model is explicitly PT-PT, though the description and verdict honestly call this out. Editorial voice is operator-grade and the weaknesses are concrete and honest.
Flags: - contextLength 4096 not explicitly verified in card excerpt — LLaMA 3.1 base supports up to 128K; verify PORTULAN's actual training/inference context before publishing - useCases includes 'portuguese-br' which contradicts the model's explicit PT-PT focus — consider removing to avoid misleading BR users
Overview
Gervásio 8B PTPT is a LLaMA 3.1 8B Instruct fine-tune from PORTULAN/University of Lisbon, trained on Portuguese-specific datasets including extraGLUE-Instruct and MMLU PT. It targets European Portuguese (PT-PT), not Brazilian Portuguese, so BR users should expect some dialect drift. MIT-licensed and commercially usable.
Strengths
- Built on LLaMA 3.1 8B Instruct — a solid, well-understood base
- Fine-tuned on curated Portuguese data: extraGLUE-Instruct, MMLU PT, Wikipedia subset
- MIT license — no commercial restrictions
Weaknesses
- Optimized for European Portuguese (PT-PT); Brazilian Portuguese (PT-BR) dialect handling is untested and likely weaker
- 4096-token context is short by current standards — long documents or multi-turn chats will hit the limit
- Only 946 HF downloads and 7 likes — very little community feedback or independent benchmarking to draw from
- 8B requires ~16 GB VRAM at full precision; quantized runs may further affect Portuguese quality
Quantization variants
Each quantization trades model quality for file size and VRAM. Q4_K_M is the most popular starting point.
| Quantization | File size | VRAM required |
|---|---|---|
| Q4_K_M | 4.4 GB | 6 GB |
Get the model
HuggingFace
Original weights
Source repository — direct quantization required.
Hardware that runs this
Cards with enough VRAM for at least one quantization of Gervásio 8B PTPT.
Models worth comparing
Same parameter band, plus what's one tier above and below — so you can decide what actually fits your hardware.
Frequently asked
What's the minimum VRAM to run Gervásio 8B PTPT?
Can I use Gervásio 8B PTPT commercially?
What's the context length of Gervásio 8B PTPT?
Source: huggingface.co/PORTULAN/gervasio-8b-portuguese-ptpt-decoder
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.
Related — keep moving
Verify Gervásio 8B PTPT runs on your specific hardware before committing money.