Complex Reasoning
Multi-step logic, mathematical proofs, and intricate problem-solving.
Qwen3.5-397B-A17B — Flagship Model for Maximum Reasoning
Qwen3.5-397B-A17B is the flagship Qwen 3.5 model — maximum reasoning power for the most demanding tasks. Try it free.
Ready To Chat
Online
ThinkingQwen3.5-397B-A17B is ready
Qwen3.5-397B-A17B is the default model for this page. Flagship large-scale Qwen 3.5 model for demanding reasoning and long-form tasks.
Pick a model, decide whether this needs web search or thinking, then start with a real prompt.
Flagship
MoE
Starter prompts
Free to use on this site via OpenRouter.
Total Params
397B
Active Params
17B
Context
262K native
License
Apache 2.0
Jump to another Qwen model
Overview
The Top of the Qwen 3.5 Lineup
Qwen3.5-397B-A17B is the largest public open-weight release in the Qwen 3.5 line. Its published scores are strong on reasoning, coding, search-agent, and multilingual tasks, while its hosted sibling Qwen3.5-Plus adds a 1M default context window and extra production features.
Current Public Scores
This model posts 87.8 on MMLU-Pro, 88.4 on GPQA, and 83.6 on LiveCodeBench v6.
Massive Expert Pool
397B total parameters provide an enormous knowledge base for expert routing.
Open Weight
Full Apache 2.0 license — download, fine-tune, and deploy without restrictions.
Qwen3.5-397B-A17B Benchmark
How Qwen3.5-397B-A17B compares to nearby models in the Qwen family.
Model
MMLU-Pro
GPQA / GPQA-family
LiveCodeBench v6
Qwen3.5-122B-A10B
Mid-tier MoE model for deeper reasoning and agent tasks.
Updated 2026-04-02
MMLU-Pro
86.7
GPQA / GPQA-family
86.6
LiveCodeBench v6
78.9
Qwen3.5-397B-A17B
Flagship open-weight Qwen3.5 model, also the base model behind Qwen3.5-Plus.
Updated 2026-04-02
MMLU-Pro
87.8
GPQA / GPQA-family
88.4
LiveCodeBench v6
83.6
Qwen3.5-Plus
HostedHosted version built on Qwen3.5-397B-A17B with additional tooling and a 1M context window.
Scores reference the Qwen3.5-397B-A17B base model.
Updated 2026-04-02
MMLU-Pro
87.8
GPQA / GPQA-family
88.4
LiveCodeBench v6
83.6
Scores are from public model cards and the qwen.ai release page. Hosted models are labeled with their open-weight base.
Updated 2026-04-02Use Cases
What Qwen3.5-397B-A17B Is Best For
Use the flagship when you need the highest quality output and reasoning depth.
Expert Coding
Architecture design, complex refactoring, and multi-repo reasoning.
Scientific Analysis
Process and reason about research papers, datasets, and technical content.
Strategic Planning
Business analysis, decision frameworks, and comprehensive project planning.
Creative Writing
High-quality long-form creative content with nuance and depth.
Benchmark Tasks
When you need the absolute best score on evaluation benchmarks.
FAQ
Qwen3.5-397B-A17B FAQ
Common questions about the flagship model.
Is 397B the best Qwen 3.5 model?
It is the largest public open-weight Qwen3.5 release and posts the strongest published benchmark scores among the open model cards on this site. Smaller models can still be faster, cheaper, or easier to operate.
Can I run it locally?
It requires significant infrastructure — multiple high-end GPUs with large VRAM. Most users access it through cloud APIs or services like OpenRouter.
How does it compare to closed-source models?
Qwen's published benchmark table places it alongside frontier closed-source models across reasoning, agent, and multilingual evaluations. The exact gap still depends on which benchmark and workflow you care about.
When should I use Plus instead?
Qwen3.5-Plus offers strong all-around performance with simpler deployment. Use 397B when you specifically need the highest reasoning ceiling.
How much VRAM does 397B-A17B need?
The full model requires 200+ GB in BF16. Quantized versions can run on 4-8 GPU setups. Cloud deployment via vLLM is the most common approach.
Is 397B-A17B the same as Qwen3.5-Plus?
Qwen3.5-Plus is the hosted version built on 397B-A17B. Plus adds production tooling and a 1M context window. The 397B-A17B weights are the open-weight version you can download.
When is 397B-A17B overkill?
For simple Q&A, drafting, or template generation, smaller models like 9B or 27B are faster and cheaper. Use 397B-A17B when you need the deepest reasoning or best code quality.
Does 397B-A17B support tool calling?
Yes. All Qwen 3.5 models support function calling. The 397B-A17B checkpoint is especially strong when tool use depends on long multi-step reasoning.