benchmarks
Qwen3-30B-A3B is Alibaba's open-weight, text-only mixture-of-experts model with 30.5 billion total parameters and 3.3 billion activated parameters, supporting both thinking and non-thinking modes. Separate Instruct-2507 and Thinking-2507 editions followed in July 2025; this profile covers the original hybrid checkpoint.
Benchmark results
Every figure links to where it was published. Compare like with like — harnesses and prompt setups differ between vendors.
| Benchmark | Score | Source |
|---|---|---|
| GPQA DiamondPhD-level science questions | 61.7%30 Aug 2026 | View |
| OTIS Mock AIME 2024-25Competition mathematics | 62.8%28 Aug 2026 | View |
Weighted mean of 2 verified benchmark results: GPQA Diamond 61.7%, OTIS Mock AIME 2024-25 62.8%.
Capabilities
- Switches between thinking and non-thinking modes within the same checkpoint.
- Supports multilingual text generation across 119 languages and dialects.
- Handles mathematical reasoning, programming and instruction-following tasks.
- Supports tool calling and agent workflows through integrations such as Qwen-Agent.
- Provides downloadable Apache-2.0-licensed weights for local deployment and customization.
What improved
- Compared with Qwen2.5, combines thinking and non-thinking behavior in one model.
- Expands multilingual coverage from 29 to 119 languages and dialects.
- The Qwen3 family increases pretraining data from approximately 18 trillion to 36 trillion tokens.
- Introduces thinking-budget control to balance reasoning depth against inference cost.
Known flaws & limitations
- The original checkpoint is text-only and does not natively process images or audio.
- The 131,072-token context requires YaRN configuration; native context is 32,768 tokens.
- Thinking mode generates additional reasoning tokens, increasing latency and token consumption.
- The model card warns that greedy decoding in thinking mode can degrade performance and cause endless repetitions.
- Static YaRN scaling can reduce performance on shorter inputs; the model card recommends enabling it only when needed.
Risks, safety & security issues
- Not yet confirmed by public reporting.
Good fit for
- Self-hosted multilingual text assistants.
- Mathematical reasoning and coding assistance with configurable thinking behavior.
- Tool-using assistants built with an appropriate agent framework.
- Applications that need both fast responses and deeper reasoning from one checkpoint.
Poor fit for
- Native image, video or audio understanding.
- Long-document workloads exceeding 32,768 tokens without context-extension configuration.
- Strict low-latency workloads with unrestricted thinking enabled.
- Deployments that require reliable thinking-mode generation with greedy decoding.
References
Profile last reviewed 10/24/2026 and re-checked automatically every few hours against the vendor's published documentation, papers and public leaderboards. Scores come from the cited benchmark results above — confirm limits and pricing against the provider's own documentation before building on them.