Qwen3.5 is Alibaba's February 2026 generation and the efficiency benchmark of the open-weight field: the 397B-A17B flagship activates just 17B parameters per token yet beats Alibaba's own trillion-parameter Qwen3-Max — at 8.6× the throughput at 32K context and 19× at 256K. All open-weight variants ship under Apache 2.0, the most enterprise-friendly license in common use.
The release is a full family, not one model: from the 397B flagship down to a 9B variant with a native 256K context window that runs on consumer hardware. That spread — one architecture, Apache 2.0 everywhere, sizes from phone to datacenter — is why Qwen has become the default base-model family for fine-tuners and the most-adopted open-weight line by download count.
Qwen3.5 specifications
| Vendor | Alibaba (Qwen team) (China) |
|---|---|
| Architecture | Mixture-of-Experts — 256 experts, 8 routed + 1 shared |
| Total parameters | 397B |
| Active parameters | 17B |
| Context window | 262K tokens (native) |
| Modalities | text |
| License | Apache 2.0 — permissive, patent-granting, free for commercial use |
| Release date | 2026-02-16 |
| Weights | Hugging Face · GitHub |
Benchmarks and reported results
| Benchmark | Result | Note |
|---|---|---|
| Throughput vs Qwen3-Max (32K context) | 8.6× faster | |
| Throughput vs Qwen3-Max (256K context) | 19× faster |
Running Qwen3.5 locally
397B total / 17B active: 4-bit builds land near 220GB for the flagship. The family is the real story locally — Qwen3.5 ships small variants down to 9B (256K context) that run on a single consumer GPU or a phone-class NPU.
What Qwen3.5 is best for
- Enterprise deployments that require Apache 2.0 licensing
- Efficiency-critical serving (17B active parameters, extreme throughput)
- Fine-tuning projects — the broadest size ladder in open weights
Frequently asked questions
What license does Qwen3.5 use?
Apache 2.0 across all open-weight Qwen3.5 models — permissive, patent-granting, and free for commercial use. The hosted Qwen3.5-Plus/Max variants on Alibaba Cloud are separate proprietary services.
Can I run Qwen3.5 on consumer hardware?
The 397B flagship no (roughly 220GB at 4-bit), but the family scales down: the 9B variant with its native 256K context runs comfortably on a single consumer GPU, and mid-size variants cover the space between.
Why is Qwen3.5 considered so efficient?
Its MoE routing activates only 17B of 397B parameters per token, and the generation was optimized for serving: Alibaba reports 8.6× the throughput of its own Qwen3-Max at 32K context and 19× at 256K — while scoring higher on quality benchmarks.