The practical open-weight pick: Qwen3.8-27B is Apache 2.0 licensed and small enough for one high-end GPU, so prompts never leave your machine. Frontier closed models still answer better.
What it is
Qwen is the model family from Alibaba's Qwen team, published on Hugging Face. The Qwen3.8 generation arrived in August 2026: Qwen3.8-27B (August 5) is a dense model under the Apache 2.0 license and the most downloaded Qwen3.8 checkpoint, alongside Qwen3.8-Flash-Next and a 2.4-trillion-parameter mixture-of-experts flagship (Qwen3.8-2.4T-A95B) that needs datacenter hardware. Licenses differ by checkpoint: the 27B is plain Apache 2.0, while the larger models ship under Qwen's own license files. Quantized to 4 bits, the 27B fits a single 24 GB graphics card and runs in tools like Ollama and llama.cpp. Checked November 12, 2026 on Hugging Face.
Best for
- Running a capable model locally or on your own servers
- Private data that cannot go to a cloud API
- Fine-tuning on your own data
What we like
- Apache 2.0 on the 27B: commercial use without extra terms
- Fits one 24 GB GPU when quantized
- Large community, many fine-tunes and quantized builds
What to watch
- Below the closed frontier models (Gemini, Claude, GPT) on public leaderboards
- Larger Qwen3.8 checkpoints use custom licenses and need multi-GPU servers
- You run and secure the infrastructure yourself
- Some companies restrict models from Chinese labs by policy
Where Qwen ranks
Alternatives and head-to-heads
Work at Qwen? Claim the listing and put this badge on your site.Claim Qwen