Qwen3.5-35B-A3B

Qwen3.5-35B-A3B is the headline model of the Qwen 3.5 Medium Series. It is a Mixture-of-Experts model with 35 billion total parameters, 256 experts, and only 3 billion parameters active per forward pass. Despite that extreme sparsity, it surpasses both Qwen3-235B-A22B and Qwen3-VL-235B-A22B - the models that were Alibaba's flagships just six months ago.

The numbers tell the story better than any marketing copy. A model activating 3 billion parameters - the compute budget of a small local model - now posts MMLU-Pro 85.3, GPQA Diamond 84.2, and TAU2-Bench 81.2 for agent tasks. That last number is a 22.7-point jump over the 235B model's 58.5 on the same benchmark.

Key Specifications

The architecture uses a repeating 4-layer cycle: three Gated DeltaNet layers (linear attention, O(n) scaling with sequence length) followed by one full softmax attention layer (quadratic but providing global context). This hybrid approach enables efficient long-context processing while maintaining the quality that full attention provides.

Benchmark Performance

The agent and vision benchmarks are where the 35B-A3B makes its strongest case. TAU2-Bench at 81.2 and AndroidWorld at 71.1 are both best-in-class for this parameter range. On ScreenSpot Pro (GUI grounding), it scores 68.6 versus Claude Sonnet 4.5's 36.2.

The coding benchmarks (SWE-bench 69.2, LiveCodeBench 74.6) are competitive but trail GPT-5-mini and the dense 27B sibling. For pure coding performance, the 27B may be the better choice in the lineup.

Key Capabilities

The 35B-A3B is a natively multimodal model - it processes text, images, and video through the same architecture rather than bolting on a vision adapter. This means it handles document understanding (OCRBench 91.0), GUI interaction (ScreenSpot Pro 68.6), and video comprehension (VideoMME 86.6 with subtitles) in a single model.

With 256 experts and only 8 routed per token plus 1 shared, the inference compute is remarkably low. At 3B active parameters, aggressive quantization (4-bit GPTQ/AWQ) can bring VRAM requirements down to consumer GPU territory. Multi-step Token Prediction (MTP) training enables speculative decoding for faster inference.

The 262K native context extends to approximately 1M tokens via YaRN rope scaling, and the 3:1 DeltaNet-to-attention ratio means most of that context is processed with linear complexity rather than quadratic.

Pricing and Availability

The model is Apache 2.0 licensed and available for download on HuggingFace and ModelScope. There is no API pricing - it is a self-hosted open-weight model.

For those who want a managed API, Qwen3.5-Flash is the production-aligned hosted version at $0.10/M input tokens. The model is also available for testing at Qwen Chat.

Quantized versions are available on HuggingFace for lower VRAM requirements.

Strengths

• 3B active parameters achieve frontier-class results - exceptional compute efficiency
• Best-in-class agent benchmarks (TAU2-Bench 81.2, AndroidWorld 71.1) at this size
• Native multimodal with strong vision (MMMU 81.4, MathVision 83.9, OCR 91.0)
• Apache 2.0 license with 201 language support
• Consumer GPU viable with quantization

Weaknesses

• SWE-bench (69.2) and LiveCodeBench (74.6) trail GPT-5-mini and the dense 27B sibling
• 256-expert MoE can be sensitive to quantization quality
• Full weight matrix (35B) must stay in memory even though only 3B activates
• Gated DeltaNet framework support still maturing in vLLM/SGLang/TensorRT-LLM
• Self-reported benchmarks - independent evaluations pending

Related Coverage

• Qwen 3.5 Medium Series Drops Four Models
• Alibaba's Qwen 3.5 Claims to Beat GPT-5.2
• Qwen 3 Review
• Open Source LLM Leaderboard
• Coding Benchmarks Leaderboard
• Open Source vs Proprietary AI Guide

Sources:

• Qwen3.5-35B-A3B Model Card (HuggingFace)
• Qwen Team on X - Medium Series announcement
• Qwen3.5: Towards Native Multimodal Agents (Blog)
• Qwen3.5 Developer Guide (NxCode)