What is Qwen3.6?
Qwen3.6 was the first open-weight release in Alibaba's Qwen3.6 generation, published on Hugging Face in April 2026 after the February Qwen3.5 series. It comes in two shapes that share a design and a benchmark table: Qwen3.6-27B, a dense model where all 27 billion parameters run on every token, and Qwen3.6-35B-A3B, a mixture-of-experts model with 35 billion parameters but only about 3 billion active at a time.
Both are causal language models with a vision encoder, so they read text, images and video and write text. Native context is 262,144 tokens, extendable to about 1,010,000 with YaRN scaling, which the card warns can hurt short-prompt quality. The generation focuses on agentic coding and repository-level reasoning, and adds a preserve_thinking option that keeps reasoning from earlier turns to cut repeated work. The weights are Apache 2.0.
Is Qwen3.6 worth it?
The headline of Alibaba's own table is that the dense Qwen3.6-27B is the better model. It scores above Qwen3.6-35B-A3B on nearly every row: 77.2 against 73.4 on SWE-bench Verified, 53.5 against 49.5 on SWE-bench Pro, and 48.2 against 28.7 on the SkillsBench average, where it even edges past the 45.3 Alibaba reports for Claude 4.5 Opus. The 35B-A3B wins only a handful of vision tests, such as RealWorldQA at 85.3 and OCRBench at 90.0. The reason to pick it anyway is speed: with 3 billion parameters active per token it runs far faster than a dense 27B, which matters on a busy server.
Read the numbers with the usual caution. Alibaba ran every benchmark itself, and several, QwenWebBench, QwenClawBench, Claw-Eval, are its own. Against Claude 4.5 Opus the picture is mixed rather than a win: the 27B ties Opus at 59.3 on Terminal-Bench 2.0 and edges it on SkillsBench and QwenClawBench, but Opus still leads every SWE-bench variant and the knowledge tests, scoring 30.8 against 24.0 on Humanity's Last Exam.
The catch: this is last generation. Alibaba's own Qwen3.8-27B, published in August 2026 under the same Apache licence, reports higher coding scores and is the one to reach for first. Qwen3.6 matters now as a proven, freely licensed fallback and as the model many hosts and tools were built around. On reasoning and knowledge it was never more than a small step up from Qwen3.5-27B.
Our take: if you are choosing between the two sizes here, the dense 27B is the one to run for quality and the 35B-A3B only if you need the throughput; if you are choosing a generation, Qwen3.8-27B has already passed both.
Where it earns its place
A developer running a local coding agent on a workstation GPU
Get repository-level coding help without a per-token bill
Run the dense Qwen3.6-27B, which tops its own generation on SWE-bench Verified at 77.2, in vLLM or llama.cpp. The preserve_thinking option keeps earlier reasoning so iterative edits reuse the KV cache instead of rethinking each turn.
A platform team serving an internal assistant to hundreds of staff
Host a permissively licensed model that answers fast under load
Serve Qwen3.6-35B-A3B, whose 3 billion active parameters per token give far higher throughput than a dense 27B on the same card, accepting that it scores a little below the dense model on most tests.
A SaaS team adding image and document understanding
Read screenshots, charts and short videos alongside text
Either size takes image and video input; the 27B scores 82.9 on MMMU and 87.7 on VideoMME. Apache 2.0 means no revenue or user caps to track when you self-host.
Before you commit
- Setup
- Weights for both sizes on Hugging Face and ModelScope, compatible with Transformers, vLLM, SGLang and KTransformers. Thinking is on by default; set enable_thinking false to turn it off, and preserve_thinking to keep prior reasoning. Qwen Chat hosts a free version.
- Integrations
- OpenAI-compatible serving through vLLM and SGLang; image and video input; function calling and structured output through the standard Qwen tooling.
- Data and privacy
- Apache 2.0: commercial use with no revenue caps or acceptable-use annex when self-hosted. Hosted use on Qwen Chat follows Alibaba's own terms, which the model card does not restate.
- Cost at scale
- Self-hosted, the cost is GPU memory and compute: the dense 27B computes every parameter per token, while the 35B-A3B activates about 3 billion, so the MoE serves more requests per card at a small quality cost.
- Limits
- 262K native context, about 1M with YaRN, which the card says may degrade short inputs; all published benchmarks are Alibaba's own; the newer Qwen3.8-27B reports better coding scores.
What’s good
- Apache 2.0 weights for both the dense and MoE versions
- Dense 27B beats the 35B-A3B on almost every benchmark
- Reads images and video as well as text
- 262K native context, preserve_thinking cuts repeated reasoning
What’s not
- Superseded by Qwen3.8-27B under the same licence
- All benchmarks run by Alibaba, several on in-house tests
- Trails Claude 4.5 Opus on SWE-bench and knowledge tests
- 35B-A3B justifies itself only on speed, not quality
Best for
Teams who want a proven, permissively licensed coding model and are choosing between a dense build for quality and an MoE build for throughput.
Qwen3.6 alternatives
Qwen3.6: common questions
- Is Qwen3.6 worth it?
- Two strong Apache-licensed coders where the surprise is that the smaller dense 27B beats its larger mixture-of-experts sibling on almost every test. We rate it 4.2 out of 5.
- How much does Qwen3.6 cost?
- Free weights under Apache 2.0 for both sizes. Hosted free on Qwen Chat; you pay the GPU to self-host or a third-party host's per-token rate.
- What are the downsides of Qwen3.6?
- Superseded by Qwen3.8-27B under the same licence; all benchmarks run by Alibaba, several on in-house tests; trails Claude 4.5 Opus on SWE-bench and knowledge tests; 35B-A3B justifies itself only on speed, not quality.
- Who is Qwen3.6 best for?
- Teams who want a proven, permissively licensed coding model and are choosing between a dense build for quality and an MoE build for throughput.
- How much does Qwen3.6 cost for a team or in production?
- Self-hosted, the cost is GPU memory and compute: the dense 27B computes every parameter per token, while the 35B-A3B activates about 3 billion, so the MoE serves more requests per card at a small quality cost.
- Is Qwen3.6 safe for company data?
- Apache 2.0: commercial use with no revenue caps or acceptable-use annex when self-hosted. Hosted use on Qwen Chat follows Alibaba's own terms, which the model card does not restate.
- What are the best Qwen3.6 alternatives?
- The closest alternatives we have reviewed are Qwen3.8-27B, Gemma 4 and Muse Glimmer.
Answers checked Nov 14, 2026.
Make Qwen3.6? Show this rating on your site
