Qwen3.5
Alibaba·
LLMsopen-weightcloud + localMulti-GPU Self-Host
Overview
Major architectural upgrade over Qwen3. 397B MoE (17B active) flagship with 1M context, natively multimodal, 8-19x higher decoding throughput. Covers 200+ languages.
Capabilities and innovations
397B total / 17B active MoE (flagship)1M context windowNatively multimodalReasoning by default200+ languagesApache 2.08-19x Throughput vs Qwen3-MaxNative Multimodal All Sizes122B-A10B Runs on MacBook 64GB
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| GPQA Diamond | 72% | unsourced |
| Humanity’s Last Exam (no tools) | 18% | unsourced |
| SWE-bench Verified | 48.5% | unsourced |
| MMLU | 90.2% | unsourced |
| MMLU-Pro | 87.18% | unsourced |
- SWE-bench Pro: not reported by the vendor
- Terminal-Bench 2.x: not reported by the vendor
- MMMU-Pro: not reported by the vendor
Architecture and hardware
- Parameters
- 397B total, 17B active per token (MoE)
- Estimated VRAM at Q4
- ~229 GB, Multi-GPU Self-Host class
Reliability
- Hallucination rate (Vectara HHEM)
- 10.7% (lower is better)independent
- Agentic tool use (τ-bench)
- 77.5% (higher is better)third party
Links
More from Alibaba
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.