Qwen /
Qwen3.6-35B-A3B
A 35B open-weight model with a vision encoder, built mainly for coding and agent-style tasks.
- 3.0k likes
- 4.0M downloads a month
- 36B parameters
- Advanced
- Vision & language
Qwen3.6-35B-A3B is the first open-weight model in the Qwen3.6 series, released after Qwen3.5 in February. It is a causal language model with a vision encoder. It uses a mixture-of-experts design: 256 experts, with 8 routed experts plus 1 shared expert used for each token. That is why only 3B of the 35B parameters are active at once. The default context length is 262,144 tokens, and the card says it can be extended up to 1,010,000.
The card says this release focuses on agentic coding and adds an option called Thinking Preservation, which keeps reasoning from earlier messages in a conversation. Honest limits: the benchmark scores come from the Qwen team, not from independent testing. It does not win everywhere, and in the table I could see, Qwen3.5-27B beats it on some tests, such as SWE-bench Pro and Tool Decathlon. The part of the card I was given does not say how much memory or what hardware you need to run it, so check that before downloading.
What you need to run it
36B parameters. The weights take about 72 GB at full 16-bit precision (fits a 128 GB Mac or a GPU with 128 GB), or about 22 GB as a 4-bit download, which runs on a 32 GB Mac or a GPU with 32 GB. Long inputs and images need more memory on top.
What it does well
Strong coding-agent scores for its size
On SWE-bench Verified it scores 73.4, against 70.0 for Qwen3.5-35B-A3B and 17.4 for Gemma4-26B-A4B in the card's own table. On Terminal-Bench 2.0 it scores 51.5, against 40.5 for the earlier Qwen3.5-35B-A3B.
Only 3B parameters active at a time
It has 35B parameters in total, but a mixture-of-experts design uses just 3B per token. The card does not turn this into a speed or memory figure, so test it on your own setup.
Very long context
It handles 262,144 tokens natively, and the card says it can be extended to 1,010,000 tokens. That helps when you want to give it a lot of code or documents at once.
Thinking Preservation
A new option keeps reasoning context from earlier messages. The card says this streamlines iterative development and reduces overhead.
Reasoning and knowledge results
It scores 86.0 on GPQA and 85.2 on MMLU-Pro. On MMLU-Pro it is roughly level with the other models in the table rather than clearly ahead.
What you could build
- Ask it to work on a frontend task, such as building or fixing a web page component.
- Point it at a small repository and ask it to explain how the code fits together, or to suggest a bug fix.
- Build a coding assistant or agent that works through a task step by step and keeps its earlier reasoning between turns.
- Give it a screenshot or image together with a question and get a text answer back.
- Paste in a long document or several files and ask questions that need the whole context.
Get started
The card says the weights work with Hugging Face Transformers, vLLM, SGLang and KTransformers, and you can try the model online at chat.qwen.ai (Qwen Chat). The part of the card I was given has no install or run commands, so follow the full model page or the Qwen blog post for exact steps.
Commands are copied from the model card. Read it on Hugging Face for the full setup.
More vision & language models
google/gemma-4-31B-it
Google DeepMind's 31B open model that reads text and images, reasons, writes code, and calls tools.
Qwen/Qwen3.6-27B
A 27B open-weight model from Qwen that reads text and images, built mainly for coding work.
microsoft/Florence-2-large
A small vision model that captions, detects objects and reads text from images using prompts.
Qwen/Qwen2.5-VL-7B-Instruct
A 7B model that reads images, charts, documents and video, and answers in text.