Qwen3.6-35B-A3B
Qwen3.6-35B-A3B is a small mixture-of-experts model from Alibaba's April 2026 Qwen3.6 release, under Apache 2.0. It has about 36B total and 3B active parameters across 256 experts, a hybrid linear/full attention stack and a vision encoder. Thinking and non-thinking modes are both supported.
Specifications
- Context window
- 256Kabout 197,000 words, or 390 pages
- Max output
- Not published
- Input price
- $0.10per 1M tokens
- Output price
- $0.95per 1M tokens
- Cached input
- $0.10per 1M tokens
- Weights
- OpenApache-2.0
- Parameters
- 36B total, 3B activeMoE
- Input types
- Text, Images, Video
- Output types
- Text
- Reasoning mode
- Yes
- Tool calling
- Yes
- Released
- Apr 16, 2026Current
DeepInfra serverless price (256K context); DeepInfra lists the cached-input rate at the same $0.10. Official context is 262,144 native, extensible to about 1M with YaRN.
What it costs in practice
| Workload | Monthly cost at list price |
|---|---|
| Support chatbot30,000 conversations a month, 2,000 input and 400 output tokens each | $17 |
| Document summaries2,000 documents a month, 12,000 input and 800 output tokens each | $3.92 |
| Coding agent1,000 tasks a month, 60,000 input and 6,000 output tokens each | $12 |
Standard rates, no caching or batch discounts. Run your own numbers.
Strengths
- 3B active parameters give fast token generation
- Apache 2.0 with image and video input
- Tool calling supported in vLLM and SGLang; Qwen recommends Qwen Code and Qwen-Agent
Trade-offs to weigh
- All ~36B parameters must be held in memory
- Small active size limits depth on hard reasoning tasks
Where it fits
Alibaba (Qwen) positions Qwen3.6-35B-A3B for:
Where you can use it
- Hugging Face
- Ollama
- DeepInfra
Run it yourself
Released under Apache-2.0. Approximate memory needed just to load the weights:
| Precision | Memory for weights |
|---|---|
| 16-bit (BF16/FP16) | 72 GB |
| 8-bit (FP8/INT8) | 36 GB |
| 4-bit (Q4_K_M, about 4.85 bits) | 22 GB |
This is a mixture-of-experts model: all 36B total parameters must sit in memory, but only about 3B are used per token, so it runs faster than a dense model of the same size.
Add memory for the KV cache (it grows with context length and concurrent users) and 1-2 GB of runtime overhead. Size it precisely in the VRAM calculator.
Alternatives to consider
| Model | Tier | Context | Input / output per 1M | |
|---|---|---|---|---|
| Ministral 3 8BMistral AI | Compact | 256K | $0.15 / $0.15 | Compare |
| Gemma 4 26B A4BGoogle | Compact | 256K | $0.07 / $0.34 | Compare |
| gpt-oss-20bOpenAI | Compact | 128K | $0.075 / $0.30 | Compare |
| Granite 4.2 8BIBM | Compact | 128K | $0.06 / $0.25 | Compare |
Frequently asked questions
How much does Qwen3.6-35B-A3B cost?
Qwen3.6-35B-A3B lists at $0.10 per million input tokens and $0.95 per million output tokens, with cached input at $0.10 (checked October 10, 2026). DeepInfra serverless price (256K context); DeepInfra lists the cached-input rate at the same $0.10. Official context is 262,144 native, extensible to about 1M with YaRN. Confirm current rates on the provider's pricing page before you commit.
What is the context window of Qwen3.6-35B-A3B?
Qwen3.6-35B-A3B accepts up to 262,144 tokens of context (about 197,000 words, or 390 pages). Long prompts cost more and can be slower, so most apps send only what the task needs.
Is Qwen3.6-35B-A3B open source?
Qwen3.6-35B-A3B is an open-weight model released under the Apache-2.0 license. You can download the weights from Hugging Face (Qwen/Qwen3.6-35B-A3B) and run or fine-tune them yourself, within the license terms. Open weights isn't the same as fully open source: training data and code may not be published.
Can I run Qwen3.6-35B-A3B locally?
Yes, if you have enough memory. The weights alone need about 22 GB at 4-bit quantization and about 72 GB at 16-bit, plus room for the KV cache. Use the VRAM calculator on LLModel.com to check a specific GPU or Mac.
What is Qwen3.6-35B-A3B best for?
Alibaba (Qwen) positions Qwen3.6-35B-A3B for self-hosting, low latency, agents, coding. It accepts text, images, video as input, has a reasoning (thinking) mode and supports tool calling. Test it on a sample of your own tasks before you choose it for production.
Sources
- huggingface.co/Qwen/Qwen3.6-35B-A3B
- deepinfra.com/pricing
- github.com/QwenLM/Qwen3.8
- huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/main/config.json
- huggingface.co/api/models/Qwen/Qwen3.6-35B-A3B
- ollama.com/library
Facts checked against these pages on October 10, 2026. Spot an error? Tell us.
Still choosing a model?
Send us your use case and constraints. We'll come back with a shortlist and an evaluation plan you can run on your own data.
- Free and vendor-neutral: no model provider pays for placement.
- Your details go only to partners you agree to.