Qwen3.5-35B-A3B

Qwen3.5 35B A3B

Quick answer: Qwen3.5 35B A3B is Alibaba's September 2025 compact MoE scoring 61.0% BrowseComp, 47.4% HLE, 85.3% MMLU-Pro, 70.2% IFBench, and 75.1% MMMU-Pro. Apache 2.0 — 35B total / 3B active.

At a Glance

Where Qwen3.5 35B A3B leads

• 47.4% HLE — strong for 3B active params
• 85.3% MMLU-Pro — excellent academic breadth
• 61.0% BrowseComp — competitive web research
• Apache 2.0 — fully open
• 3B active params: very low inference cost

Where it lags

• No SWE-Bench or coding-specific data
• September 2025: pre-dates Qwen3.6 generation

Best for: Cost-efficient MoE open-source deployments needing HLE + MMLU-Pro coverage; compact inference with Qwen3.5-tier reasoning.

What Qwen3.5 35B A3B Is

Qwen3.5 35B A3B is the compact MoE variant of Qwen3.5, running 3B active parameters from 35B total. Released September 2025 alongside the dense Qwen3.5 27B, it trades some flexibility for lower inference cost.

With 47.4% HLE and 85.3% MMLU-Pro at 3B active params, it outperforms many larger dense models on these benchmarks.

Specifications

Pricing

Open weights under Apache 2.0 — self-host at no cost.

Public Benchmark Scores

Qwen3.5 35B A3B vs Alternatives

Qwen3.5 35B A3B vs Qwen3.5 27B: virtually equal HLE (47.4% vs 48.5%) but 35B A3B runs at 3B active — 9x lower inference cost. For production cost-efficiency: 35B A3B.

Related Models & Benchmarks

• Compare: Qwen3.5 27B · Qwen3.6 35B A3B · Qwen3 30B A3B
• Benchmarks: BrowseComp · Humanity's Last Exam · MMLU-Pro

Frequently Asked Questions

Specs from Alibaba's Qwen3.5 35B A3B release (September 2025) and Benchgen evaluations. Last updated 2026-07-24.