0.375
2.25
Chat
Active
Qwen3.6-35B-A3B
Alibaba's newest open-source mixture-of-experts model activates just 3 billion parameters at inference time, yet it goes toe-to-toe with dense models four to nine times its active size on agentic coding, reasoning, and multimodal tasks.
Qwen3.6-35B-A3B is the latest open-source release from Alibaba's Qwen team, following the proprietary Qwen3.6-Plus launch. It's built on a sparse Mixture-of-Experts architecture, meaning only a small slice of the model is "awake" for any given token, which dramatically reduces compute without sacrificing output quality.
With just 3B active parameters, Qwen3.6-35B-A3B handles inference at the cost of a sub-4B dense model. You get the reasoning depth of something much larger without the GPU budget to match.
The model was explicitly trained and evaluated for multi-step coding agents, tool use, and MCP server interactions. It's not just a chat model — it's designed to plan, execute, and iterate inside automated pipelines.
Vision is a first-class citizen, not an add-on. The model processes images, diagrams, documents, and video frames natively, and its visual reasoning benchmarks are surprisingly strong for a model of this active-parameter count.
Switch between extended chain-of-thought reasoning (thinking mode) and fast, direct responses (non-thinking mode) within the same model. Both modes are supported via a simple API flag — no separate model download needed.
The MoE design isn't just a numbers game. The way Qwen3.6-35B-A3B routes information through its expert layers is what separates it from earlier sparse models that struggled with coherence and reasoning depth.
Whether you're building a coding agent, processing technical documents, or running long autonomous tasks, Qwen3.6-35B-A3B covers ground that typically requires much heavier models.
Resolve GitHub issues, navigate codebases, run tests, and iterate — without a human in the loop. SWE-bench scores confirm it can actually land patches, not just suggest them.
Terminal-Bench 2.0 is arguably the most realistic CLI benchmark available. Qwen3.6-35B-A3B leads the pack here with a 51.5 score — 10 points ahead of the next-best model in its class.
Best-in-class MCPMark score (37.0) means it reliably selects, calls, and interprets tool outputs inside MCP-based agent frameworks — a critical capability for production automation.
OmniDocBench and CC-OCR scores show it can accurately parse and reason about complex PDFs, tables, charts, and scanned documents — not just plain text.
Top-tier VideoMMMU and MLVU scores make it viable for summarizing lecture recordings, analyzing surveillance footage, or processing instructional video content at scale.
AIME 2026 at 92.7 and GPQA at 86.0 — it handles undergraduate-to-olympiad-level math and science questions with a reliability that matters when you're building serious research tooling.
The Mixture-of-Experts architecture breaks the model's feed-forward layers into many independent "expert" sub-networks. For any given input token, a learned routing mechanism selects only a small subset of those experts to activate. The rest sit idle. The result is that the total parameter count, which determines model capacity and knowledge, is 35B, but the compute required per forward pass is equivalent to a much smaller dense model.
Roughly speaking, Qwen3.6-35B-A3B costs about the same as running a 3B dense model in terms of FLOPs per token. In practice, you'll need enough VRAM to hold all 35B parameters in memory (~70GB in BF16, or ~40GB in 4-bit quantization), but the throughput and latency are comparable to running a much smaller model. Big memory footprint, small compute per inference.
Yes, the weights are released as open-source checkpoints. Standard fine-tuning approaches work, though MoE models have some quirks around expert collapse and routing stability. Tools like LLaMA-Factory and Axolotl have added MoE support. For most use cases, LoRA adapters targeting the attention layers work well without touching the expert routing.
Qwen3.6-35B-A3B is the latest open-source release from Alibaba's Qwen team, following the proprietary Qwen3.6-Plus launch. It's built on a sparse Mixture-of-Experts architecture, meaning only a small slice of the model is "awake" for any given token, which dramatically reduces compute without sacrificing output quality.
With just 3B active parameters, Qwen3.6-35B-A3B handles inference at the cost of a sub-4B dense model. You get the reasoning depth of something much larger without the GPU budget to match.
The model was explicitly trained and evaluated for multi-step coding agents, tool use, and MCP server interactions. It's not just a chat model — it's designed to plan, execute, and iterate inside automated pipelines.
Vision is a first-class citizen, not an add-on. The model processes images, diagrams, documents, and video frames natively, and its visual reasoning benchmarks are surprisingly strong for a model of this active-parameter count.
Switch between extended chain-of-thought reasoning (thinking mode) and fast, direct responses (non-thinking mode) within the same model. Both modes are supported via a simple API flag — no separate model download needed.
The MoE design isn't just a numbers game. The way Qwen3.6-35B-A3B routes information through its expert layers is what separates it from earlier sparse models that struggled with coherence and reasoning depth.
Whether you're building a coding agent, processing technical documents, or running long autonomous tasks, Qwen3.6-35B-A3B covers ground that typically requires much heavier models.
Resolve GitHub issues, navigate codebases, run tests, and iterate — without a human in the loop. SWE-bench scores confirm it can actually land patches, not just suggest them.
Terminal-Bench 2.0 is arguably the most realistic CLI benchmark available. Qwen3.6-35B-A3B leads the pack here with a 51.5 score — 10 points ahead of the next-best model in its class.
Best-in-class MCPMark score (37.0) means it reliably selects, calls, and interprets tool outputs inside MCP-based agent frameworks — a critical capability for production automation.
OmniDocBench and CC-OCR scores show it can accurately parse and reason about complex PDFs, tables, charts, and scanned documents — not just plain text.
Top-tier VideoMMMU and MLVU scores make it viable for summarizing lecture recordings, analyzing surveillance footage, or processing instructional video content at scale.
AIME 2026 at 92.7 and GPQA at 86.0 — it handles undergraduate-to-olympiad-level math and science questions with a reliability that matters when you're building serious research tooling.
The Mixture-of-Experts architecture breaks the model's feed-forward layers into many independent "expert" sub-networks. For any given input token, a learned routing mechanism selects only a small subset of those experts to activate. The rest sit idle. The result is that the total parameter count, which determines model capacity and knowledge, is 35B, but the compute required per forward pass is equivalent to a much smaller dense model.
Roughly speaking, Qwen3.6-35B-A3B costs about the same as running a 3B dense model in terms of FLOPs per token. In practice, you'll need enough VRAM to hold all 35B parameters in memory (~70GB in BF16, or ~40GB in 4-bit quantization), but the throughput and latency are comparable to running a much smaller model. Big memory footprint, small compute per inference.
Yes, the weights are released as open-source checkpoints. Standard fine-tuning approaches work, though MoE models have some quirks around expert collapse and routing stability. Tools like LLaMA-Factory and Axolotl have added MoE support. For most use cases, LoRA adapters targeting the attention layers work well without touching the expert routing.
.webp)
.png){kind=small}
.png){kind=small}
