Qwen3.8-27B Ships Open Weights with 262K Context

Qwen published the Qwen3.8-27B model card and open weights on Hugging Face under Apache 2.0. It is a dense 27-billion-parameter model, not a mixture-of-experts build, and includes a vision encoder for image and video understanding.

The release follows Qwen3.8-Max, which Qwen announced on 2 August. The 27B checkpoint is a separate open-weight release for teams that want a smaller model they can inspect and serve themselves.

Qwen lists a native context length of 262,144 tokens and documents extension to 1,000,000 tokens with RoPE scaling. That larger figure is a deployment option, not the native context limit. The model accepts text, images and video and returns text.

Thinking is enabled by default. Qwen exposes xhigh, medium and low reasoning effort, with xhigh as the default, and also allows thinking to be switched off. Preserved thinking is enabled by default so earlier reasoning blocks can remain available during multi-turn agent work.

Qwen reports gains over Qwen3.6-27B across its selected coding, professional-work and agent evaluations. Its model card lists 61.7 on SWE-bench Pro, 73.0 on Terminal Bench 2.1 and 70.7 on CoWorkBench.

Those are maker-reported results, not independent validation. Some comparisons use Qwen's own harnesses or private benchmarks, and the exact serving configuration affects long-context and agent performance. Teams should test the checkpoint on their own repositories and workflows before replacing a production model.

The full checkpoint is available from Qwen's official Hugging Face organisation. Qwen says its managed Qwen Cloud version will add a one-million-token default context and built-in tools, but the model card still labels that service as coming soon.

OpenRouter already lists the route qwen/qwen3.8-27b. Its current metadata shows a 262,144-token context, a 131,072-token maximum output, and pricing of $0.45 per million input tokens and $3.20 per million output tokens. OpenRouter does not currently confirm tool or function-calling support for this route.

The final Qwen model-card update was recorded at 15:00:01 UTC on 14 August, or 17:00:01 SAST. OpenRouter's provider metadata carries a created time of 15:55:10 UTC, exactly 55 minutes and 9 seconds later.

AIMI first observed the OpenRouter route at 01:34:42 UTC on 15 August, or 03:34:42 SAST. That was 9 hours, 39 minutes and 32 seconds after OpenRouter's metadata time and 10 hours, 34 minutes and 41 seconds after the Qwen model-card update. These timestamps describe model-card publication, provider metadata and endpoint observation. They are not interchangeable release dates.

AZ Labs does not sell a Qwen subscription plan. The relevant product is the AZ Labs AI Gateway, which gives applications one integration layer for model routing, usage visibility and failover across supported providers.

A gateway route still needs task-level testing. Context limits, reasoning controls, tool schemas and pricing can differ between providers even when they serve the same canonical model.