A 27B Local Model Just Topped the Leaderboard — Qwen3.8-27B Is Open Source

On August 14, 2026, the Qwen team at Alibaba open-sourced Qwen3.8-27B: a 27B-parameter dense model, Apache-2.0 licensed.

Then the benchmark table dropped. SWE-bench Pro: 61.7 — first place.

Look at who's in that table: Qwen3.6-27B, Qwen3.7-Plus, Muse Glimmer-30B, and Opus4.6 Max — Anthropic's flagship cloud model.

A 27B model you can run locally just pushed a cloud flagship off the top of the leaderboard. The official repo puts it plainly: the 27B dense model outperforms the larger Qwen3.7-Plus overall.

DeepSWE went 3x in one generation

The most striking cell in the table is DeepSWE 1.1 (agentic coding):

Qwen3.6-27B scored 13.3. Qwen3.8-27B scores 42.2. A 3x jump, one generation apart.

The same pattern repeats across the board:

• QwenSWEBench: 49.3 → 79.0
• OSWorld (computer use): 63.9 → 84.3 — first place
• WebArena (browser use): 48.8 → 64.8 — first place
• LiveCodeBench v6: 83.9 → 90.3 — first place
• CoWorkBench (long-horizon office work): 61.0 → 70.7 — first place
• Agents' Last Exam: Pass@1 10.6 → 20.4, Score 27.3 → 42.9 — first place

The gains cluster around one capability: agentic work — planning, reacting to environment feedback, carrying multi-step tasks to completion.

The architecture trick: 3 linear attention layers per 1 full attention

Attention was never about quality. It's about cost. Full attention scales O(N²) with sequence length. Push context to 100K or 1M tokens and compute explodes.

Qwen3.8-27B's answer is a simple ratio: 64 layers, grouped in blocks of 4 — three Gated DeltaNet (linear attention) layers for every one Gated Attention (full attention) layer, repeated 16 times.

3:1.

Linear attention layers absorb long sequences and vision tokens cheaply. Full attention layers keep precision for hard reasoning. You don't pay the quadratic tax on every block.

The architecture debuted in Qwen3.5 — Qwen's from-scratch, natively multimodal foundation model, trained on interleaved text, image, and video tokens. Qwen3.8 builds on that base, adds multi-token prediction (MTP), and ships 262,144 native context, extensible to 1,000,000 via YaRN.

Thinking mode you can actually control

A genuinely useful design choice: the model thinks by default, but you decide.

• enable_thinking: False drops the <think>...</think> block per request — direct answers when you want them.
• reasoning_effort tunes depth: xhigh, medium, low. Depth vs. cost is your call.
• preserve_thinking (default on) keeps reasoning context from previous turns, so multi-turn conversations don't forget how they got somewhere.

For product builders this beats the usual all-or-nothing thinking toggle: chat mode off, hard problems on xhigh, same model.

Computer, browser, phone — the agent wave is the story

The multimodal agent scores tell the same story:

• OSWorld-Verified (computer use): 84.3, up from 63.9
• WebArena-Verified (browser use): 64.8, up from 48.8
• AndroidWorld (mobile use): 81.9, first place
• RecreationBench (app recreation across desktop/mobile/web): 47.1, up from 29.8
• SWE-MM (multimodal software engineering): 38.6, up from 25.7
• Vision2Web (visual web development): 62.9, up from 45.0

Read those numbers right: the model isn't just chatting — it operates real interfaces. Desktops, browsers, phone screens are all fair game. That's the native multimodal training at work: text, image, and video together, hour-long videos included.

How to run it

Three paths:

Self-hosted: vLLM has an official recipe, SGLang has a cookbook, TokenSpeed is supported, and Hugging Face Transformers works natively.

API: Qwen Cloud's hosted version is coming, with 1M context by default and built-in tools.

Edge: MediaTek announced Day-0 adaptation for the Dimensity cockpit platform C-X1 and flagship mobile chips on release day. On a 64-core XuanTie C950 target configuration, decode runs at 30+ tokens/s with a 1.9-second TTFT.

Two gotchas: full-frame video input currently works via vLLM (--media-io-kwargs '{"video": {"num_frames": -1}}'), and when calling Qwen Cloud, disable thinking with "enable_thinking": False rather than the local chat_template_kwargs spelling.

The honest caveats

Before the hype takes over: HLE is 30.8 vs. Opus4.6 Max's 40.0, and GPQA still goes to the cloud flagship. Nobody should claim a 27B dense model dethrones frontier cloud models across the board.

But "a local 27B that touches the edge of cloud flagships" would have been unthinkable a year ago. Now it's a download away.

Three things this release settles:

Dense small models are punching up. First place on multiple agentic leaderboards, officially stated to beat Qwen's own larger Qwen3.7-Plus. The local-deployment value proposition finally has flagship-grade scores behind it.

Architecture dividends are paying out. Hybrid linear attention solves the long-context cost problem; 262K native with 1M extension keeps open models in the context arms race.

The open ecosystem is compounding. 460+ Qwen models open-sourced, Apache-2.0, full framework support, silicon vendors adapting on day zero.

The model is on Hugging Face: Qwen/Qwen3.8-27B. One vLLM command and it's serving.