Teams wanting an open coding model that runs on one server will find Qwen3.6-27B beats Alibaba's far larger previous flagship on terminal work. It trails Muse Glimmer at tool orchestration and answers slowly, so new projects should compare its newer sibling before committing.
What changed
- · Release Qwen3.6-27B launches, beating its own 397B predecessor on coding source
Qwen3.6-27B is Alibaba's open-weight dense model from April 2026, a 27-billion-parameter system that reads text, images and video. Its Hugging Face card reports 53.5 on SWE-bench Pro, and Alibaba's own newer sibling has since overtaken it on the vendor's tables.
Where it sits
- 58 tok/stokens/sOutput speedHigher is better#42 / 56peer median 95 tok/scited: Artificial Analysis
- 77.2%% solvedSWE-bench VerifiedHigher is better#19 / 39peer median 76.5%per source, see benchmark scores
- 87.8%% correctGPQA DiamondHigher is better#32 / 63peer median 87.8%per source, see benchmark scores
Mid-pack on SWE-bench Verified (rank 19 of 39), and one of 88 whose vendor states it does not train on customer data.
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
Provider: Alibaba Cloud (Qwen Team) · Family: Qwen
More about Alibaba Cloud (Qwen Team) on Hokai
Context window: 262,144 tokens · Max output: 131,072
Input modalities: text, image, video · Output: text, tool-calls
About Qwen3.6-27B
Qwen3.6-27B is a 27-billion-parameter dense model from the Qwen Team at Alibaba Cloud, released in April 2026 under Apache 2.0 as a leaner open-weight step down from the Qwen3.5 mixture-of-experts flagship. Alibaba followed it in August 2026 with Qwen3.8-27B, and Artificial Analysis now labels this checkpoint deprecated for its own benchmarking and points readers to the newer one.
On the Hugging Face model card, Alibaba reports 77.2 on SWE-bench Verified, 53.5 on SWE-bench Pro and 59.3 on Terminal-Bench 2.0, ahead of its 397-billion-parameter Qwen3.5-397B-A17B on the same table (76.2, 50.9 and 52.5). Reasoning rows read 87.8 on GPQA Diamond, 94.1 on AIME 2026, 86.2 on MMLU-Pro and 83.9 on LiveCodeBench v6. The larger Qwen3.5 model still leads on GPQA Diamond and MMLU-Pro. Every figure here is vendor-reported on Alibaba's own scaffolds, not a Hokai measurement.
Under the hood, 64 layers repeat a block of three Gated DeltaNet layers and one gated-attention layer, which keeps long-context memory lower than full attention. Alibaba says static YaRN scaling stretches the native window to 1,010,000 tokens while warning that this can hurt short prompts. Input covers text, images and video, output is text and tool calls, and an optional thinking-preservation mode carries earlier reasoning forward across turns.
The weights are free on Hugging Face, which also hosts an FP8 variant, and managed hosting exists through OpenRouter, Fireworks AI and NVIDIA NIM. Artificial Analysis measures 58.1 output tokens per second, below the 82 it lists as its average.
Where the alternatives win: Meta's comparison for Muse Glimmer puts Qwen3.6-27B behind on MCP-Atlas and DeepSearch QA, so broad tool orchestration favours Glimmer. Gemma 4 31B is the other open model in Alibaba's table, but it trails on its coding rows. Anyone starting a new deployment should compare Qwen3.8-27B first, since Alibaba's own side-by-side lists 90.3 against 83.9 on LiveCodeBench v6 for the newer model. The local coding models guide and the guide to running AI locally cover self-hosting trade-offs, and how to choose the right LLM covers the wider shortlist.
Pricing
The weights cost nothing to download, so self-hosting with vLLM or SGLang (both documented on the model card) costs GPU time and electricity only. Hokai found no first-party per-token rate for this checkpoint in the sources read. Third-party hosting is billed per token: OpenRouter's public models API listed $0.45 per million input tokens and $2.70 per million output tokens on 2026-10-28, up from the $0.289 and $2.40 recorded in August, and Artificial Analysis shows $0.60 and $3.60.
Alibaba Cloud sells its newer sibling on a separate managed API, and the cheapest LLM API guide compares hosts.
| Tier | Rate |
|---|---|
| Self-hosted (Apache 2.0) | Free weights from Hugging Face; cost is hardware and electricity only. |
| OpenRouter (hosted) | Third-party pay-per-token hosting, no GPU required. Rate read from OpenRouter's public models API on 2026-10-28. |
Key Features
- 27B Dense Coding & Reasoning Model: A dense transformer that matches or beats Alibaba's much larger 397B-parameter Qwen3.5 mixture-of-experts flagship on coding rows of the model card, at a fraction of the parameter count.
- Strong Reasoning Scores: Reports 94.1 on AIME 2026 on Alibaba's table, ahead of Gemma's 89.2 and the larger Qwen3.5's 93.3.
- Long Context Window: Handles long codebases in one pass, with a documented YaRN scaling path to much longer inputs than its native window.
- Thinking Preservation: Multi-turn conversations keep their prior reasoning steps in context automatically, so follow-up turns in a long coding or research session don't need the model to re-derive earlier logic from scratch.
- Native Image and Video Input: Accepts text, image, and video inputs natively, broader multimodal coverage than most same-size open reasoning models, though output remains text-only.
Pros
- Leads sustained coding-agent work among open 27B-class models on the vendor's table, and edges Muse Glimmer on SWE-bench Verified in Meta's own comparison.
- Free permissively licensed weights, with hosted options on Fireworks AI, NVIDIA NIM and OpenRouter and an FP8 checkpoint on Hugging Face.
- Efficient dense architecture delivers flagship-level coding results without the inference cost of a much larger mixture-of-experts model.
Cons
- Trails Muse Glimmer on Meta's two headline agentic-orchestration benchmarks: 62.5 vs 75.5 on MCP-Atlas and 71.1 vs 74.6 on DeepSearch QA.
- No native audio input or output; voice applications need a separate ASR/TTS pipeline.
- Slower than typical for its class on inference throughput, trading some speed for its reasoning and coding accuracy.
Benchmarks
- IFBench: 69.1% vendor-reported · 28 Oct 2026 — How precisely the model follows detailed instructions, % passing.
- MMLU-Pro: 86.2% vendor-reported · 28 Oct 2026 — A harder version of the 57-subject knowledge exam, % correct.
- MMMU-Pro: 75.8% vendor-reported · 28 Oct 2026 — College-level questions that need reading images and diagrams, % correct.
- AIME 2026: 94.1% vendor-reported · 28 Oct 2026 — Competition-level maths problems from the 2026 exam, % solved.
- MCP Atlas: 62.5% independent · 10 Aug 2026 — Tasks solved by calling external tools over the Model Context Protocol, % solved.
- GPQA Diamond: 87.8% vendor-reported · 28 Oct 2026 — PhD-level science questions that are hard to search for, % correct.
- SWE-bench Pro: 53.5% vendor-reported · 28 Oct 2026 — Harder, longer real-repository coding tasks, % solved.
- LiveCodeBench v6: 83.9% vendor-reported · 28 Oct 2026 — Recently published coding problems, % solved.
- SWE-bench Verified: 77.2% vendor-reported · 28 Oct 2026 — Real GitHub issues fixed end to end, % solved.
- Terminal-Bench 2.1: 63.4% vendor-reported · 28 Oct 2026 — Multi-step tasks completed in a real command line, % solved.
- Humanity's Last Exam: 24% vendor-reported · 28 Oct 2026 — Expert-written questions across many fields, % correct.
- Output speed: 58 tok/s cited: Artificial Analysis · 28 Oct 2026 — Median tokens written per second as measured by Artificial Analysis.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
How much does Qwen3.6-27B cost to run?
Self-hosting costs hardware and electricity only. OpenRouter, the main third-party host, was charging $0.45 for input and $2.70 for output per million tokens at the last check, versus $0.289 and $2.40 in August.
How does Qwen3.6-27B compare to Muse Glimmer and Gemma4-31B?
On Alibaba's table it scores 77.2 on SWE-bench Verified against Gemma4-31B's 52.0, and 87.8 against 84.3 on GPQA Diamond. Meta's Muse Glimmer comparison puts it behind on MCP-Atlas (62.5 vs 75.5) and DeepSearch QA (71.1 vs 74.6) while ahead on SWE-bench Verified (77.2 vs 76.0).
Can Qwen3.6-27B be used commercially under its license?
Yes. The Hugging Face card lists the Apache 2.0 license, which permits commercial use, modification and redistribution. The weights are downloadable, so a team can run the model without calling any Alibaba service.
How long a prompt does Qwen3.6-27B accept, and how quickly does it answer?
The native window is 262,144 tokens, and the card says YaRN scaling extends it to 1,010,000 while cautioning that static scaling can hurt short texts and advising at least 128K to preserve thinking. Artificial Analysis clocks 58.1 output tokens per second against a listed average of 82.
Who should use Qwen3.6-27B, and who should pick Qwen3.8-27B instead?
It suits teams already running coding agents or terminal automation on this checkpoint and wanting open weights without a mixture-of-experts footprint. New projects should evaluate Qwen3.8-27B first, which Alibaba's side-by-side lists at 89.2 against 87.8 on GPQA Diamond and 73.0 against 63.4 on Terminal Bench 2.1 (Terminus).
Hokai guides that mention Qwen3.6-27B
- Hugging Face vs OpenPipe: Which Should You Use in 2026?: OpenPipe stopped training on 30 July 2026. We compare Hugging Face TRL, CoreWeave Forge, Together AI, Fireworks and Overmind on price, RL support and data.