Model releases · 11 August 2026
The most widely downloaded open model line currently available, released across a size ladder that runs from something that fits on a phone to a 397-billion-parameter sparse model — under the same permissive licence at every rung.
| Released by | Alibaba (Qwen team) |
|---|---|
| Dense sizes | 0.8B, 2B, 4B, 9B, 27B |
| Sparse sizes | 35B-A3B, 122B-A10B, 397B-A17B |
| Context length | 262,144 tokens native, extensible to about 1,010,000 |
| Vocabulary | 248,320 tokens |
| Licence | Apache 2.0 across every size |
| Weight format | safetensors; FP8 builds published for the larger sparse sizes |
| Most downloaded | Qwen3.5-9B, over 12 million downloads |
| Hosted at | Qwen on Hugging Face |
Qwen3.5 is a family rather than a model. Five dense sizes and three Mixture-of-Experts sizes, all under Apache 2.0, sharing a 248,320-token vocabulary and a 262,144-token native context window that the model card says extends to roughly 1,010,000 tokens.
The download figures put the family's reach in perspective: Qwen3.5-9B alone has passed 12 million downloads, and four other sizes are past 2 million each.
| Size | Type | Typical use |
|---|---|---|
| 0.8B | Dense | Phones and embedded; fits comfortably under a gigabyte quantised |
| 2B | Dense | On-device assistants; the base others build French, Japanese and other language specialists on |
| 4B | Dense | The largest size that runs comfortably on current phones |
| 9B | Dense | Single consumer GPU; the family's most downloaded size |
| 27B | Dense | 24 GB cards at 4-bit |
| 35B-A3B | MoE | 3B active per token — MoE quality at small-model inference cost |
| 122B-A10B | MoE | Workstation and multi-GPU |
| 397B-A17B | MoE | The family's frontier size; FP8 build also published |
The 0.8B and 2B variants are the ones that show up as the base for other people's work. Kurakura AI's French Luth-2 models are built on Qwen3.5-2B; specialists in other languages follow the same pattern. A permissively licensed, well-trained base at that size is scarce, and Qwen is currently the default answer.
One detail carried down from the base: the 2B variant ships with a vision-language architecture. Downstream models inherit that architecture whether or not they were trained on vision data, which is why some derivatives explicitly warn against using them for image tasks.
The model card is unusually specific about the trade-off. Default context is 262,144 tokens, and the guidance is to keep at least 128K even when memory is tight, because Qwen3.5 uses extended context for its reasoning behaviour — cutting the window too far degrades more than just recall.
Qwen has announced Qwen3.8, with a 27B size named publicly and a larger Max variant flagged for open release. Weights were not published at the time of writing.
| Runtime | Supported | Notes |
|---|---|---|
| llama.cpp | Yes | Widely supported; GGUF conversions exist for every size from third parties. |
| MLX | Yes | Community MLX conversions available across the size range. |
| transformers | Yes | Reference implementation from the lab. |
| vLLM | Yes | The model card documents vLLM serving commands, including tensor-parallel setups for the long-context configuration. |
| Official quantised build | Partial | FP8 builds are published for the larger sparse sizes; smaller sizes rely on community quantisation. |
Five dense sizes — 0.8B, 2B, 4B, 9B and 27B — and three Mixture-of-Experts sizes: 35B-A3B, 122B-A10B and 397B-A17B. All are Apache 2.0.
Apache 2.0, at every size in the family, including the largest sparse models.
262,144 tokens by default, extensible to approximately 1,010,000. The lab advises keeping at least 128K because the model uses extended context for its reasoning behaviour.
The 0.8B, 2B and 4B dense variants are the ones in the range that current phones can hold quantised, with 4B being the practical ceiling on recent hardware.
No. A 27B size and a larger Max variant have been announced publicly, but the weights had not been published at the time of writing.
OnDevice LLM is a private AI assistant that runs entirely on your iPhone — no account, no cloud, and nothing you type leaves the device.
Published 11 August 2026