What is Qwen3.6-27B?
Qwen3.6-27B is the dense member of Alibaba's Qwen3.6 open family: all 27B parameters work on every token, so it serves without mixture-of-experts routing. It reads text, images and video, thinks before it answers by default, and can carry its earlier reasoning across turns in agent work. Alibaba reports it beats its own 397B Qwen3.5 flagship on every major coding test it ran.
It is still on Hugging Face and Alibaba's API, but Qwen has since released Qwen3.8-27B, a model of the same size and design with higher scores and a lower API price.
Is Qwen3.6-27B still worth using after Qwen3.8-27B?
For new work, mostly no: Qwen3.8-27B is the same size and design, scores higher in Alibaba's tests and costs less on Alibaba's API.
- Same size and design
- Both are 27B dense models with the same layer layout, so hardware that runs one runs the other.
- Coding scores
- Alibaba's table gives Qwen3.8-27B 61.7% on SWE-bench Pro against 53.5% for 3.6, and 73.0 against 63.4 on Terminal Bench 2.1.
- API price and context
- Qwen3.8-27B costs $0.50 / $3 per million tokens against $0.60 / $3.60, and takes up to 1M tokens per request instead of 256K.
- Still available
- Qwen3.6-27B remains on Hugging Face and Alibaba's API, so setups already tuned for it keep working.
Who is Qwen3.6-27B best for?
Qwen3.6-27B is best for developers who run coding agents on their own hardware, plug a model into Claude Code, Qwen Code or OpenClaw, or need a coder that reads screenshots and video.
Self-hosted coding agents
Alibaba reports it beats its far larger Qwen3.5-397B-A17B on SWE-bench Verified, SWE-bench Pro and Terminal-Bench 2.0, in a model small enough to host yourself.
Coding tools you already use
It works with Claude Code, Qwen Code and OpenClaw, through Model Studio's OpenAI- and Anthropic-compatible API or your own server.
Work from screenshots and video
One checkpoint reads images and video as well as text, in thinking or non-thinking mode, so it can work from screenshots, diagrams and documents.
Simpler self-hosting
It is a dense model rather than a mixture of experts, so serving it needs no expert routing setup, and the Apache 2.0 licence keeps your code on your own hardware.
What's included with Qwen3.6-27B?
Thinking you can switch off, reasoning kept across turns, image and video input, OpenAI- and Anthropic-style API access, and official FP8 weights.
- Thinking on or off
- It reasons before answering by default, one request setting turns that off, and both modes cost the same on Alibaba's API.
- Preserved reasoning
- The preserve_thinking option keeps reasoning from earlier turns, which Qwen recommends for agent tasks.
- Image and video input
- The same weights handle documents, charts, visual questions and phone-screen agent tasks, with no separate vision model.
- OpenAI- and Anthropic-style API
- Alibaba's Model Studio accepts both formats, so Claude Code, Qwen Code and OpenClaw connect with a base URL and a key.
- Official FP8 weights
- Qwen publishes an FP8 build of about 31 GB next to the 56 GB BF16 weights.
- Common serving stacks
- It runs in SGLang, vLLM, KTransformers and Hugging Face Transformers, with multi-token prediction for faster output.
How much does Qwen3.6-27B cost?
Qwen3.6-27B is free to download and run under the Apache 2.0 licence, and through Alibaba Cloud Model Studio's API it costs $0.60 per million input tokens and $3.60 per million output tokens, with thinking on or off.
- Apache 2.0 licence, on Hugging Face and ModelScope
- Run it on your own hardware
- 262,144-token context, extendable to 1,010,000
- Same price in thinking and non-thinking modes
- Up to 256K input tokens per request
- 1M free tokens for 90 days in Singapore
API prices are for Model Studio's Singapore (International) region; its US (Virginia) and Frankfurt regions do not offer Qwen3.6-27B, and Qwen Cloud, Qwen's official API service, lists the same $0.60 / $3.60. In the China (Beijing) region it costs $0.412564 per million input tokens and $2.475384 per million output tokens. The free API quota counts 90 days from Model Studio activation or the model's release, whichever is later. Self-hosting costs only your own hardware. Full pricing on alibabacloud.com. We checked these prices on .
How do readers rate Qwen3.6-27B?
No one has rated Qwen3.6-27B yet. Be the first.
Sign in to rate. One rating per account.
Qwen3.6-27B pros and cons
It is a strong local coder that fits on one high-end GPU, but it is slow on Macs, can stumble in long tool-calling runs, and now trails Qwen3.8-27B.
Pros
- Many developers run it daily as their main local coding model, and several say it is the first local model that holds up next to Claude Code for their work.
- Quantised to 4 to 6 bits it fits on a single 24 GB to 32 GB GPU, and users report little quality loss from quantisation.
- People who tried both find it clearly smarter than the faster Qwen3.6-35B-A3B, and Alibaba's own scores agree: 77.2% against 73.4% on SWE-bench Verified.
- Apache 2.0 weights allow commercial use, and Alibaba's API charges the same $0.60 / $3.60 per million tokens with thinking on or off.
Cons
- It is slow on Macs, Strix Halo machines and 16 GB cards: users report about 5 to 8 tokens a second there, against 25 or more for Qwen3.6-35B-A3B.
- Tool calling can break down in long agent runs; several users see it repeat a failed call or fail after a few dozen tool steps.
- It still trails frontier closed models on large, loosely specified tasks, and users say it can break working code and lean on tests to catch the damage.
- Qwen3.8-27B, the newer model of the same size, scores higher in Alibaba's tests and costs less on the API, and Alibaba's US region does not offer 3.6.
Is Qwen3.6-27B open source?
Yes. Qwen3.6-27B's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Qwen3.6-27B alternatives?
Qwen3.6-27B is an open source alternative to Nemotron 3, Hunyuan Vision 1.5 Thinking and Nemotron 3 Super. GLM-4.6V, NVIDIA Nemotron 3 Ultra and Step3-VL-10B do the same job and are open source too.
Nemotron 3FreeNVIDIA's open model family for agents, reasoning and long context.
Hunyuan Vision 1.5 ThinkingPaidTencent's vision model for OCR, charts and visual reasoning on the TokenHub API.
Nemotron 3 SuperFreeNVIDIA's open 120B MoE reasoning model with a 1M context.
GLM-4.6VOpen SourceZ.ai's open vision-language family with 128K context and tool calling.
NVIDIA Nemotron 3 UltraOpen SourceNVIDIA's 550B open-weight reasoning model with a 1M context.
Step3-VL-10BOpen SourceStepFun's 10B open vision-language model that fits in 24 GB of VRAM.
Questions about Qwen3.6-27B
What is Qwen3.6-27B?
Qwen3.6-27B is Alibaba's free, open-weight 27B dense model for agentic coding that reads text, images and video, built for developers who host their own model or want a low-cost API.
Who is Qwen3.6-27B for?
Qwen3.6-27B is best for developers who run coding agents on their own hardware, plug a model into Claude Code, Qwen Code or OpenClaw, or need a coder that reads screenshots and video.
Is Qwen3.6-27B free?
Yes. Qwen3.6-27B has a free Open weights plan: apache 2.0 licence, on Hugging Face and ModelScope.
How much does Qwen3.6-27B cost?
Qwen3.6-27B's paid plans are Model Studio API at $0.60 / $3.60 per million input / output tokens.
Is Qwen3.6-27B open source?
Yes. Qwen3.6-27B's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Qwen3.6-27B alternatives?
The closest alternatives on Recaply are Nemotron 3, Hunyuan Vision 1.5 Thinking, Nemotron 3 Super, GLM-4.6V, NVIDIA Nemotron 3 Ultra and Step3-VL-10B.
Where can I try Qwen3.6-27B?
On its official website, qwen.ai.
Can I use Qwen3.6-27B commercially?
Yes. Qwen3.6-27B's weights are released under the Apache 2.0 licence on Hugging Face and ModelScope, which allows commercial use, changes and redistribution as long as you keep the licence and its notices with the model.
Can I run Qwen3.6-27B on my own computer?
Yes. Qwen3.6-27B's BF16 weights take about 56 GB and Qwen's official FP8 build about 31 GB, and it serves through vLLM, SGLang, KTransformers or Transformers. Qwen advises keeping at least 128K tokens of context so its thinking works well.
Is Qwen3.6-27B better than Qwen3.6-35B-A3B?
For coding quality, yes by Alibaba's numbers: 77.2% against 73.4% on SWE-bench Verified and 59.3 against 51.5 on Terminal-Bench 2.0. The 35B-A3B uses only 3B active parameters, so it runs faster and costs less on Alibaba's API, at $0.375 / $2.25 per million tokens.
How do I turn off thinking in Qwen3.6-27B?
Qwen3.6-27B thinks by default. On your own server, set enable_thinking to false in the chat template settings; on Alibaba's Model Studio API, set enable_thinking to false in the request itself. Both modes cost the same on Alibaba's API.
Our verdict
Qwen3.6-27B is still a capable open coding model, but new projects should start with Qwen3.8-27B.
It beats Alibaba's own 397B Qwen3.5 flagship on coding tests and runs on a single high-end GPU, yet Qwen3.8-27B keeps the same size and design, scores higher and costs less on the API.
- Skip it for
- People on Macs or Strix Halo machines who need fast replies, and anyone who needs Alibaba's API in its US region.
- Check first
- Run your own tasks against Qwen3.8-27B before you settle on 3.6.
Qwen3.8-27B
Qwen3-Coder-Next
Rnj-1
Mellum2
Ornith-1.0
MiniMax M2.1