# Qwen3.8-27B — a 27B open model that beats Opus 4.6 Max on SWE-bench Pro

> A 27B dense model with open Apache-2.0 weights that reads images and video and runs long agentic coding jobs.

- Org: Alibaba (Qwen)
- Date: 2026-08-14
- Category: model
- Source: https://huggingface.co/Qwen/Qwen3.8-27B
- Canonical page: https://ai-tldr.dev/releases/qwen-3-8-27b/

## In short

Qwen3.8-27B is Alibaba's new 27B dense open-weights model with built-in vision, released under Apache-2.0. It scores 61.7 on SWE-bench Pro and 84.3 on OSWorld-Verified, ahead of Opus 4.6 Max on both.

## What it is

Qwen3.8-27B adds native vision to Alibaba's compact dense line: it takes images and hour-scale video alongside text, and thinking mode is on by default with a per-request off switch. The weights are on Hugging Face under Apache-2.0, in both BF16 and FP8. The context window is 262,144 tokens natively and stretches to 1,000,000.

## How it works

Under the hood sit 64 layers that repeat one hybrid block 16 times: three Gated DeltaNet linear-attention layers for every single full Gated Attention layer. Qwen3.8-27B was trained with multi-token prediction, and a `reasoning_effort` setting picks low, medium or xhigh reasoning depth per call. Thinking from earlier turns is kept by default so agents do not lose their train of thought.

## Why it matters

Strong agentic coding now comes in weights you can host yourself. On SWE-bench Pro the 27B model scores 61.7 against 53.4 for Opus 4.6 Max, and computer use on OSWorld-Verified climbs from Qwen3.6-27B's 63.9 to 84.3. Teams that could not send code to a closed API get a model that competes on the same tasks.

## Who it's for

developers running local or self-hosted models

## Try it

Model id: Qwen/Qwen3.8-27B (BF16) or Qwen/Qwen3.8-27B-FP8

## Quick facts

- Maker: Alibaba (Qwen)
- Size: 27B dense, 64 layers
- Context window: 262,144 native, up to 1,000,000
- License: Apache-2.0
- Modality: Text, image and video in
- Formats: BF16 and FP8 checkpoints
- Thinking: On by default, tunable low / medium / xhigh

## FAQ

### Is Qwen3.8-27B free for commercial use?

Qwen3.8-27B ships under the Apache-2.0 licence, which allows commercial use, modification and redistribution without a separate agreement with Alibaba. That is looser than Qwen3.8-Max, the 2.4T flagship, which uses Qwen's own bespoke licence. Both the BF16 and the FP8 checkpoints on Hugging Face carry the same Apache-2.0 terms.

### How much better is Qwen3.8-27B than Qwen3.6-27B?

Qwen3.8-27B improves on Qwen3.6-27B on every benchmark Qwen published. Agentic terminal coding on Terminal Bench 2.1 rises from 63.4 to 73.0, SWE-bench Pro from 53.5 to 61.7, and computer use on OSWorld-Verified from 63.9 to 84.3. The biggest jump is DeepSWE 1.1, which goes from 13.3 to 42.2.

### How does Qwen3.8-27B score on a third-party benchmark?

Qwen3.8-27B scores 52 on the Artificial Analysis Intelligence Index, a third-party aggregate benchmark run outside Alibaba. Artificial Analysis ranks the model #1 of 135 open-weights models in its small 4B-40B size class, and lists its context window as 256k tokens. That matters because the SWE-bench Pro and OSWorld numbers come from Qwen's own model card.

### Can I get Qwen3.8-27B as a hosted API?

A hosted Qwen3.8-27B is planned but not live yet. The model card points to Qwen Cloud for managed inference and says the 27B service will offer a 1M-token context by default plus built-in tools, but marks it as coming soon with no date given. For now the open weights are the way to run it.

### Which serving frameworks support Qwen3.8-27B?

Qwen publishes ready-made recipes for Qwen3.8-27B on SGLang, vLLM and TokenSpeed, and the Hugging Face repository holds the post-trained weights in Transformers format. Qwen recommends those dedicated serving engines over plain Transformers for production or high-throughput work, and warns that throughput varies a lot between frameworks.

### Does Qwen3.8-27B beat closed frontier models?

Qwen3.8-27B leads Opus 4.6 Max on several agentic benchmarks in Qwen's own table: 61.7 versus 53.4 on SWE-bench Pro, 84.3 versus 72.7 on OSWorld-Verified, and 90.3 versus 88.8 on LiveCodeBench v6. Opus 4.6 Max still wins Terminal Bench 2.1 at 78.2 to 73.0, plus GPQA Diamond and HLE.

## Links

- [Model Card](https://huggingface.co/Qwen/Qwen3.8-27B)
- [Qwen3.8 Announcement](https://qwen.ai/blog?id=qwen3.8)
- [FP8 Weights](https://huggingface.co/Qwen/Qwen3.8-27B-FP8)
- [vLLM Recipe](https://recipes.vllm.ai/Qwen/Qwen3.8-27B)
- [Artificial Analysis](https://artificialanalysis.ai/models/qwen3-8-27b)
- [HN Discussion](https://news.ycombinator.com/item?id=49299684)