Released April 2026, Qwen3.6-27B packs Alibaba's coding and reasoning strength into a 27B dense model with a 262K-token context window, beating its own much larger MoE predecessor on coding. It edges past Muse Glimmer on SWE-bench Verified but trails on agentic orchestration benchmarks like MCP-Atlas.
Qwen3.6-27B, Alibaba's open-weight model released April 2026, scores 87.8% on GPQA Diamond and leads most 27B-class open rivals on core reasoning and coding benchmarks. It trails agentic orchestration leaders like Muse Glimmer, and ships under an open license with native image and video input support.
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 Alibaba's open-weight model in the Qwen line, released in April 2026 by the Qwen Team. It is a 27-billion-parameter dense transformer that outperforms Alibaba's own 397-billion-parameter Qwen3.5 mixture-of-experts flagship on coding benchmarks, positioned as a leaner, self-hostable alternative rather than a direct replacement for larger closed models. On its own published evaluation suite, Qwen3.6-27B scores 87.8% on GPQA Diamond, 94.1% on AIME 2026, 86.2% on MMLU Pro, and 83.9% on LiveCodeBench v6, leading most models in its weight class on core reasoning and coding tasks. Multimodal document and visual reasoning reaches 75.8% on MMMU Pro, and the model includes a thinking-preservation feature that retains full reasoning traces across multi-turn conversations instead of discarding them after each response. In Meta's own published comparison for its rival Muse Glimmer model, Qwen3.6-27B scored 62.5 on MCP-Atlas and 71.1 on DeepSearch QA, both behind Glimmer's 75.5 and 74.6, and also behind Glimmer on AIME 2026 math scoring. The one benchmark where Qwen3.6-27B comes out ahead is SWE-bench Verified, where it scores 77.2% against Glimmer's 76.0%, particularly on sustained terminal work and computer control; Google's Gemma4-31B trails both models on that same benchmark family. The model ships with a 262,144-token native context window, extendable further with YaRN scaling for exceptionally long documents, built on a hybrid gated-attention design that balances long-context recall against inference cost. Native function calling and structured output are built in, and the thinking-preservation feature persists reasoning state across turns rather than resetting it on every call. Input covers text, images, and video; output is text only, including tool calls and structured formats. Alibaba has published strong benchmark results across both English and Chinese evaluation sets, consistent with the broader Qwen lineage's multilingual focus, though there is no native audio input or output in this release. Weights are free to download and self-host under an Apache 2.0 license, so the only direct cost is compute. Third-party hosting is available on OpenRouter at a modest per-token rate, well below typical closed frontier API pricing, and Fireworks AI and NVIDIA NIM both offer hosted access alongside the official Hugging Face weights. Terminal-Bench (59.3%) rounds out Alibaba's published scores, showing a model tuned for structured coding and sustained terminal work rather than broad agentic orchestration. Alibaba has not published a dedicated safety or preparedness report specific to this release; the Hugging Face model card covers intended use, limitations, and standard safety fine-tuning. Qwen3.6-27B fits teams that want strong, self-hosted coding-agent and reasoning performance without a mixture-of-experts model's inference overhead, particularly for sustained terminal automation and computer-use workloads. Teams building broad multi-tool agentic orchestration will likely get better results from Muse Glimmer instead, and teams prioritizing raw document-reasoning throughput over coding may prefer Gemma4-31B. It arrives as the efficiency-focused successor to the Qwen3.5 line, continuing Alibaba's rapid open-weight release cadence.
Pricing
Qwen3.6-27B's own weights are free under an open license; Alibaba has not published an official first-party per-token price. On OpenRouter, the primary independently verified third-party host, the model costs $0.289 per million input tokens and $2.40 per million output tokens. Self-hosting instead costs only hardware and electricity.
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 tasks, while running at a fraction of the parameter count.
- Strong Reasoning Scores: Scores 94.1% on AIME 2026 math reasoning, among the highest results in its weight class.
- Long Context Window: Handles very long codebases and documents in a single pass, thanks to an attention design built to balance recall over extended inputs against inference cost.
- 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 on sustained coding-agent work, scoring 77.2% on SWE-bench Verified, ahead of Muse Glimmer's 76.0% in Meta's own published comparison.
- Free, permissively licensed weights with day-one support across Hugging Face, Fireworks AI, and NVIDIA NIM.
- 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
- mmlu pro: 86.2
- mmmu pro: 75.8
- aime 2026: 94.1
- mcp atlas: 62.5
- gpqa diamond: 87.8
- deepsearch qa: 71.1
- livecodebench v6: 83.9
- swe bench verified: 77.2
- terminalbench hard: 59.3
- artificial analysis speed tokens per sec: 56.9
Frequently Asked Questions
How much does Qwen3.6-27B cost to run?
Self-hosting costs nothing beyond your own hardware and electricity, since the weights are freely licensed. Going through OpenRouter instead, the main third-party host, bills $0.289 for every million input tokens and $2.40 for every million output tokens processed.
How does Qwen3.6-27B compare to Muse Glimmer and Gemma4-31B?
Qwen3.6-27B wins on sustained coding-agent work, edging past Muse Glimmer on SWE-bench Verified, and it beats both rivals on core reasoning benchmarks like GPQA Diamond and AIME 2026. It trails Muse Glimmer specifically on Meta's agentic-orchestration benchmarks, MCP-Atlas and DeepSearch QA, where Glimmer leads all three models.
Is Qwen3.6-27B open source?
Yes. Qwen3.6-27B ships under the Apache 2.0 license, a permissive license that allows commercial use, modification, and redistribution, with weights hosted on Hugging Face. This makes it fully self-hostable, independent of Alibaba's own cloud infrastructure.
What is Qwen3.6-27B's context window?
Qwen3.6-27B natively handles up to 262K tokens of context, with YaRN scaling available to stretch that further for unusually long documents. It uses a hybrid gated-attention design that keeps inference cost manageable even at extended lengths.
Who is Qwen3.6-27B best for, and who should avoid it?
It's a strong fit for teams building coding agents or sustained terminal-automation workflows who want open-weight efficiency without a massive MoE model's inference cost. For pure agentic orchestration or broad multi-tool reasoning, Muse Glimmer will likely perform better, since Qwen3.6-27B trails on those specific benchmarks.