On April 16, the Alibaba Qwen team open-sourced Qwen3.6-35B-A3B under the Apache 2.0 license with no commercial restrictions.
This model is noteworthy: it has 35B total parameters but activates only 3B during inference. It uses a mixture-of-experts (MoE) architecture with a 12:1 compute sparsity ratio — delivering performance close to a 35B dense model while running at the cost of a 3B model.
On SWE-bench Verified (a real-world GitHub issue-fixing test), it scored 73.4%. By comparison, Google's open-source Gemma 4-31B scored 52.0%, a gap of 21 percentage points.
Why the architecture matters
First, the design logic behind MoE.
In a typical dense model, every parameter is used to process each token. A 35B-parameter model requires 35B of computation per token.
MoE is different. It divides parameters into many "experts" and activates only a subset per token. Qwen3.6-35B-A3B's routing mechanism calls on only the 3B-parameter expert group for each token, while the overall model retains the knowledge capacity of 35B parameters.
The result: inference speed and memory usage are on par with a 3B model, while problem-solving ability is on par with a 35B model.
On an RTX 4090, it runs at over 120 tokens per second. A 64GB MacBook Pro M4/M5 can run it directly. After Q4_K_M quantization, it is about 21GB — small enough to fit on a single mainstream consumer GPU.
This has significant implications for local deployment. Previously, running cutting-edge open-source coding models required at least an A100 or multiple 4090s. Now a single consumer-grade GPU suffices.
Comparative benchmarks
vs. Google Gemma 4-31B:
| Benchmark | Qwen3.6-35B | Gemma 4-31B |
|---|---|---|
| SWE-bench Verified | 73.4% | 52.0% |
| Terminal-Bench 2.0 | 51.5% | 42.9% |
| MCPMark (tool calling) | 37.0% | 18.1% |
| GPQA (general reasoning) | 86.0% | 84.3% |
| AIME26 (math competition) | 92.7% | 89.2% |
On tool calling (MCPMark), Qwen3.6 more than doubles Gemma 4. This benchmark directly corresponds to performance in real agent tasks, making it even more relevant than SWE-bench.
On QwenWebBench (web tasks), it achieved 1397 ELO, a 43% improvement over the previous generation.
vs. Claude Sonnet 4.5:
According to The Decoder's evaluation, on pure coding benchmarks the two models are nearly tied, with differences within the measurement error margin. Claude still leads in assistant-style conversation and general chit-chat.
Open-source details and availability
Model weights are available on Hugging Face (Qwen/Qwen3.6-35B-A3B) and ModelScope, compatible with Transformers, vLLM (>=0.19.0), SGLang (>=0.5.10), and KTransformers.
Two modes are supported:
- Thinking mode: A reasoning-like model that performs multi-step deep thinking, preserving the reasoning process for subsequent dialogue.
- Fast mode: Direct output, suitable for conversation and lightweight tasks.
Native context window is 262,144 tokens, extendable to over 1 million with YaRN.
The API is also available through Alibaba Cloud Model Studio (named "Qwen3.6 Flash"), or you can deploy it yourself.
What this means
This is Alibaba's third step in the open-source direction. First came the open-source Qwen3 base version (Apache license), then the closed-source Qwen3.6-Plus enterprise release, and now the open-source Qwen3.6-35B-A3B, a high-performance coding-specialized version.
Each step follows a clear rhythm: first build the community and developer base, then monetize with a closed-source version, then deepen the ecosystem with an open-source version.
Where does Qwen3.6-35B-A3B's 73.4% SWE-bench score rank among open-source models? Currently, only DeepSeek V4 and this model are notable in the open-source space. Among closed-source models, Claude Sonnet 4.5 is also in this range. A locally runnable MoE model achieving this score genuinely changes the landscape.
Of course, a high SWE-bench score does not guarantee production readiness. The real-world performance of coding agents also depends on context management, error recovery, and long-flow stability — none of which are captured in benchmarks. But this is a strong starting point worth serious evaluation.
Sources: CocoLoop, Alibaba's open model Qwen3.6 leads Google's Gemma 4 across agentic coding benchmarks (The Decoder); Qwen3.6-35B-A3B: 73.4% SWE-Bench, Runs Locally (Build Fast with AI); Qwen3.6-35B-A3B Complete Review: Alibaba's Open-Source Coding Model (DEV Community)