Qwen releases Qwen3.8-27B-FP8 vision-language model

Qwen has released Qwen3.8-27B-FP8 on Hugging Face, an FP8-quantized version of the post-trained Qwen3.8-27B model. Qwen describes the quantized weights' performance as nearly identical to the original, unquantized model. The quantization uses a fine-grained FP8 method with a block size of 128.

Qwen3.8-27B is a dense, 27 billion parameter causal language model with a native vision encoder, built on the architectural foundation of the earlier Qwen3.5 series. Qwen calls it the most capable generation in the Qwen open-model family so far, following what it describes as widespread community adoption of the Qwen3.5 and Qwen3.6 series. The model has 64 transformer layers, a hidden dimension of 5120, and a native context length of 262,144 tokens that Qwen says is extensible up to 1,000,000 tokens.

Qwen positions Qwen3.8-27B as a compact, deployment-friendly model that natively understands both images and video, from STEM diagrams and documents to hour-scale videos, alongside coding, professional work, research, and long-horizon agentic tasks. It ships with flexible thinking control: a thinking mode that is on by default but can be disabled per request, reasoning depth adjustable through a reasoning_effort parameter (xhigh by default, with medium and low options), and a preserve_thinking setting, also on by default, that retains reasoning context from earlier turns in a conversation.

Qwen notes that in multi-turn agentic tasks, lowering the reasoning effort does not always cut overall task completion time: faster per-turn responses can lead to insufficient analysis and repeated retries, which can raise total latency and token use.

The model card lists several benchmark suites used to evaluate the model, including SWE-bench Pro, NL2Repo-Bench, DeepSWE 1.1, an in-house QwenSWEBench and CoWorkBench, HLE, and vision-language benchmarks such as MathVision, WebArena-Verified, RecreationBench, ClawEval-MM, Vision2Web and SWE-MM, but the retained material does not include the actual score tables. Qwen also says a separate, hosted version of Qwen3.8-27B is coming through its Qwen Cloud service, with production features such as a default 1,000,000-token context length and official built-in tools; that hosted service is described as "coming soon."

Key facts

Qwen released Qwen3.8-27B-FP8, FP8-quantized weights for the dense 27-billion-parameter Qwen3.8-27B vision-language model, on Hugging Face.
The quantization uses a fine-grained FP8 method with a block size of 128, and Qwen says performance is nearly identical to the original unquantized model.
The model has a native context length of 262,144 tokens, extensible up to 1,000,000 tokens, with 64 transformer layers and a hidden dimension of 5120.
Qwen3.8-27B is built on the Qwen3.5 architecture, adds native image and video understanding, and includes flexible thinking control via a reasoning_effort parameter (xhigh, medium, or low).
A separate hosted version on Qwen Cloud is planned, with a default 1,000,000-token context and official built-in tools, described only as "coming soon."

Why it matters

Qwen calls Qwen3.8 the most capable generation in the Qwen open-model family so far, positioning it as a substantial step up in coding, professional work, research, and long-horizon agentic tasks over the earlier Qwen3.5 and Qwen3.6 series. This release also ships FP8-quantized weights for the dense vision-language variant, making the model runnable with a smaller memory footprint while, according to Qwen, nearly matching the full-precision model's performance.

Who it affects

Developers and teams running open-weight models locally or self-hosted through frameworks such as Transformers, vLLM, SGLang, or Docker Model Runner, especially those building agentic or vision-language applications that need image or video understanding or long context windows. It is also relevant to users of Qwen's hosted inference who are watching for the upcoming Qwen Cloud version of Qwen3.8-27B.

How to use it

The FP8-quantized weights are on Hugging Face as Qwen/Qwen3.8-27B-FP8 and load through Transformers' pipeline or the AutoModelForMultimodalLM/AutoProcessor classes, or can be served with vLLM, SGLang, or Docker Model Runner, each exposing an OpenAI-compatible chat completions API. Qwen recommends temperature 1.0, top_p 0.95, top_k 20 for thinking mode, and temperature 0.7, top_p 0.80, top_k 20, presence_penalty 1.5 for instruct (non-thinking) mode. Reasoning depth is set through reasoning_effort (xhigh by default, or medium/low), and preserve_thinking is on by default to retain reasoning context across turns; both can be turned off. No pricing or license terms for the FP8 weights are given in the retained material.

How solid is it

The model card names the benchmark suites used, including SWE-bench Pro, NL2Repo-Bench, DeepSWE 1.1, in-house QwenSWEBench and CoWorkBench, HLE, and vision-language benchmarks such as MathVision, WebArena-Verified, RecreationBench, ClawEval-MM, Vision2Web, and SWE-MM, evaluated largely through the Claude Code harness with disclosed methodology notes on temperature, context window, and judge models. The retained text does not include the actual score tables, so the benchmark claims cannot be checked against numbers here, and no comparison figures against competing models are given.

Risks and caveats

The claim that the quantized weights perform "nearly identical" to the full model is Qwen's own characterization, with no numeric comparison in the retained text. The model card gives no release date, no license terms, and credits no named individual researchers or institutions, speaking throughout in the collective "we"/"Qwen." The hosted Qwen Cloud version, with its default 1,000,000-token context and built-in tools, is not yet available and is described only as "coming soon," with no pricing disclosed.

“we are pleased to introduce Qwen3.8, the most capable generation in the Qwen open-model family to date”

— Qwen, Qwen3.8-27B-FP8 model card