Alibaba Qwen3.5

On February 16, 2026, Alibaba released Qwen3.5, and the headline model is a 397 billion parameter Mixture-of-Experts giant with just 17 billion active parameters per token. Shipped under Apache 2.0 on Hugging Face, it immediately became one of the most capable open-weight models ever published, with scores on AIME 2026, LiveCodeBench, and SWE-Bench Verified that rival the top proprietary frontier models.

Qwen3.5 397B-A17B: A Big Sparse Bomb

The flagship Qwen3.5 release is an ultra-sparse Mixture-of-Experts:

• 397 billion total parameters
• 17 billion active parameters per token (roughly 4.3% activation)
• Apache 2.0 license, open weights on Hugging Face
• Part of a broader Qwen3.5 family that spans from sub-1B dense models to the 397B flagship

The sparsity ratio is the key. At 17B active, serving Qwen3.5 is closer in cost to running a 17B dense model than a 397B giant, but the model draws on the full parameter pool during expert routing. That gives it frontier-class quality at mid-tier inference cost.

Benchmarks That Put It in Frontier Territory

Qwen3.5 397B-A17B posts numbers that put it squarely in the top tier of any model, open or closed:

• AIME 2026 (competition math): 93.3%
• LiveCodeBench v6 (coding): 85.0
• SWE-Bench Verified (coding agent): 76.8%

For reference, these are benchmarks where a year ago the best closed frontier API models were scoring in the 60s and 70s. Qwen3.5 is pulling the frontier down into open-weight territory.

A Whole Family, Not Just a Flagship

Alongside the 397B-A17B flagship, Alibaba released smaller Qwen3.5 variants designed for developers who want to run them on consumer or single-GPU hardware. The lineup spans dense models from roughly 600 million parameters up through a 32B dense model, plus the sparse MoE variants at 30B-A3B and the flagship 397B-A17B. Every open-weight size in the family ships under Apache 2.0.

Where to Get It

1. Hugging Face: browse the Qwen organization for all 3.5 checkpoints.
2. ModelScope: Alibaba's own model hub for users in China.
3. Ollama: pull smaller variants for local testing.
4. Alibaba Cloud API: hosted inference for teams that do not want to self-serve.
5. GitHub: reference code and examples at QwenLM/Qwen3.5.

Why Qwen Keeps Winning on Open Weights

Alibaba has now shipped a consistent cadence of open-weight flagship releases, from Qwen2 to Qwen2.5, Qwen3, Qwen3-Next, and now Qwen3.5. Each release raises the bar on benchmark scores while keeping the Apache 2.0 license. The Qwen family has become the default base model for a huge share of fine-tuned derivatives on Hugging Face, and Qwen3.5 is the new starting point for anyone building on top of open weights in 2026.