Qwen 3.6 35B A3B
alibaba/qwen3.6-35b-a3b- Alibaba
- Family · Qwen3
- Released 16 Apr 2026
- Open weights
- Text · Image · Video → Text
- Tools · JSON · Vision
Spec & list price
Context256K
Max output236K
Input / 1M$0.10
Output / 1M$0.909× input
Scores · standing
Intelligence18.2#153/489
Coding41.9#87/196
Usage
Tokens / day84.7B▲ 30% 7d
Usage rank#35
Requests / day5.3M
Tool calls / req0.23
Per request · 30 days
Avg input12.1K0.8× median
Avg output3410.5× median
Cache hit53%0.8× median
Reasoning share53%about median
Tool-call failures1.1%0.6× median
Consumption
Peak 386.1B 16 OctLow 7.0B 6 OctRank today #35
News & trends
2 mentions this week ▼ 50%this model · videos, posts, articles · 30 days
Trending · 14 days
- Qwen3.6-35B-A3B: OpenCode 2 vs Cline vs Crush | Local Coding on Strix Halo31 views
- Five engines on an M3 Max 36 GB: Kapitsa 41 tok/s dense, oMLX/Kapitsa tie at ~123 tok/s on Qwen3.6-35B-A3B, plus Flint 40B4 upvotes
- I Tested Flint 40B A2.5B Q3 and Qwen3.6 35B A3B Q479 upvotes
News · latest from Qwen3
- Memdoor: A team of 5 Qwen3.8-27B models beats Fable 5 on LiveCodeBench Hardgithub.com · 31 Oct
- Neuron Explorer for Qwen3.5-4B: See every neuron at work, make targeted editsgithub.com · 26 Oct
- How to train your own Jev for $17Together AI · 23 Sep
Benchmarks
bar = share of the top score · Benchmark data: Artificial AnalysisSpeed · median across 11 providers, latest day
Output speed#174/33464 tok/sabout median
Fastest provider126 tok/sbest of 11 providers
First token#1560.97 sabout median
First answer token#291.1 sthinking included
#3 cheapest of 20 with a similar score
Value
Intelligence Index
Blended price per 1M tokens (log)
Family progress
Intelligence Index
- Family releases33.7
Technical details
- Tokenizer
- Qwen
- License
- Apache 2.0
- Parameters
- 36B
- Moderation
- none
- Downloads · likes
- 4.6M · 3.1K
Request parameters 19
- tools
- tool_choice
- response_format
- structured_outputs
- reasoning
- include_reasoning
- temperature
- top_p
- top_k
- min_p
+9 more
- max_tokens
- stop
- seed
- frequency_penalty
- logit_bias
- logprobs
- presence_penalty
- repetition_penalty
- top_logprobs
Interesting facts
- 1.8× the output room of similar models (235.9K against a 128.0K median).
- 2nd of 15 Qwen3 models by tokens last month.
When to use
Best for
- Agentic tool use and instruction following
- Image and video understanding on a budget
- Open-weight self-hosted assistants
- Science question answering
Think twice for
- Hardest terminal and coding agents
- Audio input
- Very large single responses
Research papers
- Weaken to Strengthen: Weakening-Induced On-Policy Self-Distillation for Large Reasoning Models28 Oct
- Building Presentation Intelligence for High-Quality Slide Generation22 Oct
- Task-Phase-Conditioned Expert Routing Is Not a Reliable Proxy for Quantization Sensitivity in MoE FFN Tensors14 Oct
- Are You Reading the Right Thing? Rethinking Internal Signals for Prompt-Injection Detection14 Oct
- AssetForge: Asset-Augmented Task Synthesis for Business Automation Agents13 Oct
- Locate, Judge, Read: Role-Specialized Tool Agents over a Knowledge-Point Index for Open-Domain Multimodal Document QA9 Oct
- CTE-Bench: Counterfactual Trace Evaluation for Stateful Software Simulators29 Sep
- Training LLMs to Verbalize Evaluation Awareness28 Sep
Social
Most watched and discussed · 90 days
- Qwen3.6-35B-A3B: OpenCode 2 vs Cline vs Crush | Local Coding on Strix HaloDolmario Local AI · 31 views · 27 Oct
Same family
Qwen3- 7.2
- 6.0
- 6.9
- 6.1
- 5.2
- 13.1
Alternatives
other vendors · similar capability- S19.5
- 20.6
- 15.6
- 20.7
Using Qwen 3.6 35B A3B at volume?
Sign in to ask for a volume discount. You will continue in the OSMEXO dashboard.