Summary
Alibaba released the open weights of Qwen3.8-27B, the open sibling of the Qwen3.8 family promised at the Aug 3 Qwen3.8-Max launch. It is a dense 27.8B vision-language model with hybrid attention (Gated DeltaNet + Gated Attention), taking image, video and text input, under Apache 2.0. Native context is 262,144 tokens, extensible to 1M via YaRN. Thinking is on by default, with reasoning_effort control (xhigh/medium/low).
Why it matters
If you are picking a base model for a local or self-hosted coding agent, this is the leading candidate this cycle: near-frontier agentic and multimodal capability in a model that fits on a single GPU, under Apache 2.0.
Technical details
| Params | 27.78B BF16 (~55.6GB) |
|---|---|
| Architecture | 64 layers; 16 x (3 x (Gated DeltaNet -> FFN) -> 1 x (Gated Attention -> FFN)); MTP trained; architecture family qwen3_5 |
| Context | 262,144 native / 1,000,000 via YaRN |
| Modality | image-text-to-text (images, video, text) |
| License | Apache-2.0 |
| Benchmarks | swe_bench_pro 61.7 terminal_bench_2_1 73 livecodebench_v6 90.3 gpqa_diamond 89.2 osworld_verified 84.3 webarena_verified 64.8 deepswe_1_1 42.2 hle 30.8 ifbench 79.5 |
| Adoption | 91,917 downloads + 9,479 likes on HF within ~1 day; 319 community quantizations; Ollama day-0 support (v0.32.12/v0.32.13 on Aug 14); OpenRouter $0.45/$3.20 per Mtok |
| Runtimes | vLLM · SGLang · transformers · Ollama |
Tags
qwenopen-weightsapache-2agenticmultimodallong-contextlocal