Qwen3.6-35B-A3B

Canonical version: Qwen3.6-35B-A3B.

Qwen3.6-35B-A3B is an open-weight LLM from the Qwen family (Alibaba Cloud), released in April 2026. Following the Qwen MoE naming convention, the model has 35B total parameters with ~3B active per token (the A3B suffix), making it runnable on consumer hardware via quantization.

Architecture

• Mixture-of-experts (MoE) design — 35B total parameters, ~3B active per token
• Open-weight release under the standard Qwen licensing terms
• Distributed in GGUF format via HuggingFace (ecosystem forks by Unsloth and others)
• Typical quantized builds (Q4_K_S) land around 21 GB, suitable for consumer Apple Silicon and mid-range GPUs

Running it locally

• Simon Willison documented running Qwen3.6-35B-A3B-UD-Q4_K_S.gguf (Unsloth Dynamic, 20.9 GB) on a MacBook Pro M5 via LM Studio
• The low active-parameter count (3B) keeps inference fast despite the larger on-disk footprint
• A good fit for the AI Expert Offloading strategy on hybrid CPU/GPU setups

Notable reception

• Simon Willison's tongue-in-cheek "pelican riding a bicycle" SVG benchmark: Qwen3.6-35B-A3B produced a bicycle frame with correct geometry and a recognizable pelican, while Claude Opus 4.7 failed to render the frame in two attempts (including thinking_level: max)
• The follow-up "flamingo on a unicycle" test also favored Qwen, which even added inline SVG comments like Sunglasses on flamingo!
• Simon explicitly cautions: the pelican test is intentionally absurd and this is not evidence that a 21 GB quantized model is broadly more capable than Anthropic's flagship; narrow benchmarks are easy to over-read

Why it matters

• Demonstrates that a small-active-parameter MoE, quantized to ~21 GB, can match or beat frontier proprietary models on specific creative/spatial tasks
• Reinforces the case for local, open-weight models as a practical complement to API-only frontier models
• Fits the broader pattern of Qwen shipping competitive open-weight releases on a fast cadence

References

• Official announcement: https://qwen.ai/blog?id=qwen3.6-35b-a3b
• Simon Willison's write-up: https://simonwillison.net/2026/Apr/16/qwen-beats-opus/
• Qwen on HuggingFace: https://huggingface.co/Qwen

Related

• Qwen
• Qwen3.6-27B
• Claude Opus 4.7
• LM Studio
• AI Open Weight Models
• AI Expert Offloading
• Large Language Models (LLMs)
• Simon Willison

About Sébastien

Ready to get to the next level?