Alibaba Releases Qwen3.8-27B-FP8, a 27B Dense Vision-Language Model with 1M-Token Context

Alibaba Ships Qwen3.8-27B-FP8

Alibaba's Qwen team has published FP8-quantized weights for Qwen3.8-27B, a 27-billion-parameter dense vision-language model that the company describes as "the most capable generation in the Qwen open-model family to date." The release, hosted on Hugging Face, uses fine-grained FP8 quantization with a block size of 128 and is compatible with Hugging Face Transformers, vLLM, SGLang, and TokenSpeed.

Qwen3.8-27B is built on the Qwen3.5 architectural foundation and is a causal language model paired with a vision encoder, giving it native image and video understanding — according to Qwen, including hour-scale video and STEM diagrams. The model has a hidden dimension of 5,120, 64 layers arranged in a Gated DeltaNet / Gated Attention hybrid layout, and was trained with multi-token prediction (MTP).

Context and Reasoning Controls

The model supports a native context window of 262,144 tokens, extensible up to 1,000,000 tokens, according to Alibaba. Thinking mode is enabled by default but can be disabled per request; reasoning depth is adjustable via a reasoning_effort parameter, and a preserve_thinking option retains reasoning context from prior turns.

Alibaba says a hosted version through Qwen Cloud is coming, which will ship with a 1M-token context window by default and built-in tools. Pricing for that hosted API has not yet been disclosed.

Benchmark Claims

Alibaba reports gains over Qwen3.6-27B and Qwen3.7-Plus across coding and agentic benchmarks, evaluated using the Claude Code harness. According to the company's published results, Qwen3.8-27B scores 61.7 on SWE-bench Pro (versus 53.5 for Qwen3.6-27B), 73.0 on Terminal-Bench 2.1, 90.3 on LiveCodeBench v6, and 89.2 on GPQA Diamond. On multimodal agentic tasks, Alibaba claims 84.3 on OSWorld-Verified and 81.9 on AndroidWorld. On the Humanity's Last Exam (HLE) benchmark, judged by GPT-4o, the model scores 30.8.

These figures come directly from Alibaba's own evaluation and have not been independently verified. The company compares Qwen3.8-27B against Qwen3.6-27B, Qwen3.7-Plus, a third-party "Muse Glimmer-30B" model, and an unnamed "Opus4.6 Max" reference point, though methodology details for cross-vendor comparisons are limited to the harness and context-window settings disclosed in Alibaba's benchmark notes.

What This Means

Qwen3.8-27B-FP8 extends Alibaba's strategy of shipping compact, dense open-weight models that pack vision-language and long-context capability into a size deployable on single high-memory GPUs, aided by the FP8 quantization. A 262K native context window extensible to 1M tokens, combined with tunable reasoning effort, targets long-horizon agentic and coding workflows rather than short chat interactions.

The benchmark claims — if they hold up under independent testing — would put a 27B open model competitive with much larger proprietary systems on agentic coding and multimodal tasks. That comparison should be treated cautiously: all reported scores are self-published by Alibaba, and the reference models used for comparison, including "Opus4.6 Max" and "Muse Glimmer-30B," are not standard, widely-benchmarked releases. Developers should validate these figures against their own workloads before relying on them for deployment decisions. Pricing and availability for the managed Qwen Cloud API remain undisclosed.