Qwen 3.8 27B
Qwen 3.8 27B
The open Qwen most people should run now: dense, Apache-licensed and sized for one GPU.
Updated
The read
Qwen 3.8 27B is a dense vision-language model published under Apache 2.0 in August, and it takes over from Qwen 3.6 as the practical open Qwen for self-hosting. It handles images as well as text, runs on a single high-memory GPU, and its coding and science scores are strong for its size. It is not a frontier model, and the independent leaderboards place it a step below the big sparse releases, but nothing about its licence will surprise your lawyers.
Where it shines
- Apache 2.0 with no user or revenue conditions
- Dense, so it is simple to serve, quantise and fine-tune
- Reads images as well as text
Worth knowing first
- A step below the large sparse open models on hard reasoning
- A 262,000-token native context, stretched to a million only with scaling tricks
Index score
82.9
Provisional. This score is read from the product's public capability, not from the one-prompt rebuild the fully reviewed entries went through, so treat it as a placing rather than a verdict.
- Rank
- 31 of 73
- Builds
- LLMs
- Category
- Open weights
- Output
- Apache-licensed weights, text and images in
- Pricing
- Free tier
- Best for
- Self-hosting and fine-tuning on a single GPU
Scores are our own editorial judgement, weighted craft 55, speed 10, control 15 and value 20.
Also worth a look
Full index- 05GLM-5.3 FlashGLM-5.3 FlashProvisionalA fast, MIT-licensed open model that human voters rank beside the big flagships.Open weightsFree tier89.6
- 06Xiaomi MiMo V2.6 ProMiMo V2.6 ProProvisionalThe strongest large open-weights model on the independent index, under MIT, for under a dollar out.Open weightsFree tier89.4
- 07DeepSeek V4.1 FlashDeepSeek V4.1 FlashProvisionalFrontier-adjacent answers for thirty cents per million input tokens, with the weights under MIT.Open weightsFree tier88.3
