Model Details
Qwen3.5 397B A17B
Quick answer: Qwen3.5 397B A17B is Alibaba's September 2025 large MoE model scoring 69.0% BrowseComp, 88.4% GPQA Diamond, 28.7% HLE, 87.8% MMLU-Pro, and 76.4% SWE-Bench Verified. Apache 2.0 with 397B total / 17B active parameters.
At a Glance
Where Qwen3.5 397B A17B leads
- 69.0% BrowseComp — above Qwen3.5 122B A10B (63.8%) on web research
- 88.4% GPQA Diamond — strong graduate-level science
- 87.8% MMLU-Pro — strong academic knowledge
- 91.3% AIME 2026 — excellent competition math
- Apache 2.0 — fully open
- 17B active params: efficient inference per token for 397B capacity
Where it lags
- 28.7% HLE — below Qwen3.5 122B A10B (47.5%) — this is unusual
- 76.4% SWE-Bench — below Kimi K2.6 (80.2%)
- Requires very large infrastructure (397B)
Best for: Alibaba open-weight deployments at maximum scale; BrowseComp-class web research; teams with large GPU clusters.
What Qwen3.5 397B A17B Is
Qwen3.5 397B A17B is the large-scale MoE model in Alibaba's Qwen3.5 generation — with 397B total parameters and 17B active. It is larger than the 122B A10B variant while using only 17B active params per token.
The lower HLE vs the smaller 122B A10B model (28.7% vs 47.5%) is notable — suggesting the 397B A17B may be better calibrated for factual/instruction tasks (higher BrowseComp: 69% vs 63.8%) but the 122B A10B was more thoroughly fine-tuned for academic reasoning.
Specifications
| Field | Value |
|---|---|
| Organization | Alibaba |
| License | Apache 2.0 |
| HuggingFace | Qwen/Qwen3.5-397B-A17B |
| Release date | September 2025 |
| Parameters | 397B total / 17B active (MoE) |
| Modality | Text only |
| Context window | 128K tokens |
Pricing
Open weights under Apache 2.0 — self-host. Available via Alibaba Cloud DashScope.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| BrowseComp | 69.0% | Benchgen evaluation | 2025-09 |
| GPQA Diamond | 88.4% | Benchgen evaluation | 2025-09 |
| AIME 2026 | 91.3% | Benchgen evaluation | 2025-09 |
| Humanity's Last Exam | 28.7% | Benchgen evaluation | 2025-09 |
| MMLU-Pro | 87.8% | Benchgen evaluation | 2025-09 |
| SWE-Bench Verified | 76.4% | Benchgen evaluation | 2025-09 |
Qwen3.5 397B A17B vs Alternatives
| Model | BrowseComp | HLE | MMLU-Pro | License | Active params |
|---|---|---|---|---|---|
| Qwen3.5 397B A17B | 69.0% | 28.7% | 87.8% | Apache 2.0 | 17B |
| Qwen3.5 122B A10B | 63.8% | 47.5% | 86.7% | Apache 2.0 | 10B |
| Kimi K2.6 | 83.2% | 36.4% | — | Apache 2.0 | MoE |
Qwen3.5 397B A17B vs 122B A10B: higher BrowseComp (69% vs 63.8%) but lower HLE (28.7% vs 47.5%) despite being larger. Use 122B A10B for reasoning/HLE tasks; 397B A17B for BrowseComp-heavy web research.
Related Models & Benchmarks
- Compare: Qwen3.5 122B A10B · Kimi K2.6 · Qwen3 235B A22B
- Benchmarks: BrowseComp · GPQA Diamond · AIME 2026
Frequently Asked Questions
What is Qwen3.5 397B A17B?
Alibaba's September 2025 large MoE model: 397B total / 17B active parameters, scoring 69.0% BrowseComp, 88.4% GPQA Diamond, 87.8% MMLU-Pro. Apache 2.0.Specs from Alibaba's Qwen3.5 397B A17B release (September 2025) and Benchgen evaluations. Last updated 2026-07-24.
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.