Alibaba's Qwen team dropped four mid-sized models on Monday under the Qwen 3.5 Medium banner: Qwen3.5-Flash, Qwen3.5-35B-A3B, Qwen3.5-122B-A10B, and Qwen3.5-27B. Weights for the open models are live on Hugging Face, with Flash available as a hosted API through Alibaba Cloud's Model Studio.
The standout is Qwen3.5-35B-A3B. It's a mixture-of-experts model with 35 billion total parameters but only 3 billion active per inference pass. Per Alibaba's own benchmarks, it outperforms the previous-generation Qwen3-235B-A22B, which activated 22 billion parameters. That's roughly a 7x improvement in compute efficiency, though the benchmarks are company-reported and haven't been independently verified yet.
Qwen3.5-Flash is the production wrapper around that same 35B-A3B architecture, tuned for agentic workflows. It ships with a 1-million-token context window and native function calling out of the box. Pricing on Model Studio starts at $0.05 per million input tokens and $0.40 per million output tokens in the international tier.
The larger 122B-A10B and 27B variants target multi-step reasoning and long-horizon planning tasks. Alibaba used a four-stage post-training pipeline involving chain-of-thought cold starts and reasoning-based reinforcement learning. The 122B model, running on just 10B active parameters, reportedly competes with much heavier dense models on logical consistency. All open-weight models ship under Apache 2.0.
These releases follow the flagship Qwen3.5-397B-A17B that launched on February 16. The medium series fills the gap Alibaba promised when it said smaller sizes were coming.
Bottom Line
Qwen3.5-35B-A3B matches or beats a 235B-parameter predecessor while activating roughly one-seventh the parameters, all under Apache 2.0.
Quick Facts
- Four models released: Flash, 35B-A3B, 122B-A10B, 27B
- Qwen3.5-35B-A3B: 3B active parameters (company-reported to outperform Qwen3-235B-A22B)
- Flash context window: 1 million tokens
- Flash API pricing: $0.05/1M input, $0.40/1M output (international tier)
- License: Apache 2.0 for open-weight models
