benchmarks
Alibaba's Qwen3-8B is an open-weight, text-only language model with approximately 8.2 billion parameters and switchable thinking and non-thinking modes. It supports multilingual conversation, coding and tool-use workflows, with a native 32,768-token context that can be extended through YaRN configuration.
Benchmark results
Every figure links to where it was published. Compare like with like — harnesses and prompt setups differ between vendors.
| Benchmark | Score | Source |
|---|---|---|
| AIMECompetition mathematics | 76%May 2025; AIME 2024, thinking mode | View |
| GPQA DiamondPhD-level science questions | 59.3%May 2025; thinking mode | View |
| MMLU-ProBroad expert knowledge and reasoning | 74%May 2025; thinking mode | View |
| OTIS Mock AIME 2024-25Competition mathematics | 56.1%27 Aug 2026 | View |
Weighted mean of 4 verified benchmark results: AIME 76%, GPQA Diamond 59.3%, MMLU-Pro 74%, OTIS Mock AIME 2024-25 56.1%.
Capabilities
- Switches between thinking and non-thinking modes within the same model.
- Supports mathematical reasoning, programming and general instruction-following tasks.
- Supports 119 languages and dialects across the Qwen3 family.
- Supports tool calling and agent workflows through integrations such as Qwen-Agent.
- Provides downloadable weights under the Apache 2.0 license for local deployment and customization.
What improved
- Combines reasoning and conventional conversational behavior rather than requiring separate model variants.
- Expands family-level language coverage from Qwen2.5's 29 languages to 119 languages and dialects.
- Adds controls for reasoning behavior and thinking budgets to balance computational cost against task difficulty.
- The vendor reports improved reasoning, instruction following and agent capabilities over Qwen2.5.
Known flaws & limitations
- The 131,072-token context requires YaRN configuration; native context is 32,768 tokens.
- Static YaRN can reduce performance on shorter inputs.
- Greedy decoding is discouraged because it can degrade performance and produce endless repetition.
- Thinking mode generates additional reasoning tokens, increasing output length and inference work.
- The model accepts text rather than native image, audio or video inputs.
Risks, safety & security issues
- Unsuitable decoding settings can produce repetitive output that consumes inference resources without completing a useful answer.
- Generated tool calls require application-side authorization and validation before they affect external systems.
- Thinking output requires deliberate separation from final answers in user-facing applications.
Good fit for
- Locally hosted multilingual assistants where downloadable weights and deployment control matter.
- Coding and mathematical assistance with adjustable reasoning effort.
- Text-based tool-use workflows with constrained permissions and validated actions.
- Domain adaptation and experimentation under a permissive open-source license.
Poor fit for
- Applications requiring native image, audio or video understanding.
- High-stakes decisions without independent verification and human oversight.
- Strict low-latency workloads with unrestricted thinking enabled.
- Long-context applications expecting 131,072-token support without configuration or workload testing.
References
Profile last reviewed 10/17/2026 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.