PhoGPT 4B
PhoGPT-4B is a 3.7B-parameter model pre-trained from scratch on 102B Vietnamese tokens, making it one of the few Vietnamese-first generative models available. Context window tops out at 8192 tokens. A separate chat-tuned variant (PhoGPT-4B-Chat) handles instruction following and conversation.
If Vietnamese is your only target language and your hardware is constrained, PhoGPT-4B is the most purpose-built open option available at this size. The 102B-token Vietnamese pretraining is a genuine differentiator over repurposed multilingual models. That said, 3.7B parameters will hurt on anything requiring nuanced reasoning, and the thin download numbers mean you're partly in uncharted territory. Hedge: worth testing against your actual task before committing; skip if you need multilingual support at all.
›Why this rating
Auto-generated rating (Opus 4.7 judge, claude-opus-4-7). Overall 9.10/10. License is explicitly BSD-3-Clause on the HF card and commercial use is correctly flagged. Params (3.7B), context (8192), vendor (VinAI), and family (mpt) all match the card and HF tags. Description is concrete, honest, and operator-voiced — explicitly notes the 102B-token Vietnamese pretraining, calls out low download counts as a real risk, and warns about reasoning ceilings at 3.7B. Use case is sharply scoped to Vietnamese text generation. Minor deployability gap: no mention of GGUF availability or VRAM expectations, and the custom_code MPT architecture can be a real friction point that isn't flagged. Overall a solid, niche-but-defensible row that clears the bar.
Flags: - No mention of custom_code/trust_remote_code requirement for MPT architecture — relevant deployability detail - No GGUF/quantization availability noted
Overview
PhoGPT-4B is a 3.7B-parameter model pre-trained from scratch on 102B Vietnamese tokens, making it one of the few Vietnamese-first generative models available. Context window tops out at 8192 tokens. A separate chat-tuned variant (PhoGPT-4B-Chat) handles instruction following and conversation.
Strengths
- Pre-trained from scratch on 102B Vietnamese tokens — not adapted from a multilingual base
- 8192-token context window is generous for a 3.7B model
- BSD-3-Clause license allows commercial use without royalty concerns
- Outperforms prior open-source Vietnamese-language models per VinAI's own evals
Weaknesses
- 3.7B parameters puts a real ceiling on reasoning and factual recall
- Vietnamese-only — not a viable option if you need any cross-lingual capability
- Low community traction (929 downloads, 21 likes) means limited third-party testing or bug reports
- Training data sourcing and composition are not publicly detailed beyond language and token count
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 | 2.0 GB | 3 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 PhoGPT 4B.
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 PhoGPT 4B?
Can I use PhoGPT 4B commercially?
What's the context length of PhoGPT 4B?
Source: huggingface.co/vinai/PhoGPT-4B
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.
Related — keep moving
Verify PhoGPT 4B runs on your specific hardware before committing money.