Qwen3.8-27B: AA Index 52 Matches GPT-5.6 Luna and 3M Downloads in Three Days
Alibaba released Qwen3.8-27B to Hugging Face on August 15 under an Apache 2.0 licence, and the model passed 3 million downloads within three days. Quantised versions for local inference appeared within hours of the weights going public.
The 27B is the smaller sibling of Qwen3.8-Max (2.4 trillion parameters, 95B active), which launched via API on August 2. Unlike the Max variant — whose downloadable weights are text-only and strip out the million-token context window — the 27B is a dense model designed from the start for local deployment with full multimodal capability.
Benchmark numbers
Independent evaluation by Artificial Analysis gives Qwen3.8-27B an Intelligence Index score of 52 at maximum reasoning, placing it level with GPT-5.6 Luna at max — a proprietary cloud-only model. That equivalence is the headline claim: a locally runnable, openly licensed model matching a frontier cloud inference offering on a composite capability index.
| Benchmark | Qwen3.8-27B | Claude Opus 4.8 (max) |
|---|---|---|
| AA Intelligence Index | 52 | 57 |
| AA Agentic Index | 51 | ~49 |
| SWE-bench Pro | 61.7% | 69.2% |
| OSWorld-Verified | 84.3% | 83.4% |
| CoWorkBench | 70.7% | 72.3% |
| LiveCodeBench v6 | 90.3% | — |
On the AA Agentic Index, which measures task completion across multi-step agent workflows, the 27B model scores 51 — above Claude Opus 4.8’s published result. Alibaba’s own benchmark table shows the 27B beating Claude Opus 4.6 Max on both SWE-bench Pro and LiveCodeBench, though Opus maintains a lead on Terminal-Bench 2.1, GPQA Diamond, and HLE.
SWE-bench Pro at 61.7% is a meaningful figure for a 27B model. The full Qwen3.8-Max at 2.4T manages 67.7% — a 6-point gap that narrows to near-nothing on several vision and office tasks.
What the 3M download number signals
Qwen3.8-27B follows a pattern Alibaba established with Qwen3.6-27B: release a dense smaller model alongside the Max-tier flagship, licence it permissively, and let the open-weights ecosystem do the distribution. The strategy works. 3 million downloads in three days is faster than any prior Qwen release at this size tier.
The quantised models — GGUF and GPTQ formats — appeared on Hugging Face within hours. A Q4 quantisation of the 27B model runs in roughly 20GB VRAM, putting it inside a single RTX 5090 with headroom. That’s the threshold where local inference on consumer hardware becomes practical.
Context window and multimodal
The 27B model supports a 62,000-token context window. That is smaller than the Qwen3.8-Max API’s million-token context but larger than most competing 27B-class models at the time of release. Input modalities include text and vision; audio input is not supported in the base weights release.
The Max model’s open-weights caveat
Alibaba did release Qwen3.8-Max weights the same week, but the downloadable version has limitations the API version does not: text-only input, no million-token context, and reduced reasoning-effort configuration options. For developers who need multimodal or long-context capabilities in a locally hosted model, the 27B is currently the more practical option despite its smaller parameter count.
Weights for both models are available at Qwen/Qwen3.8-27B and Qwen/Qwen3.8-Max on Hugging Face. Integration with mainstream agent frameworks (LangGraph, AutoGen, CrewAI) and IDE plugins is documented on the Qwen AI Platform.