Smaller models can excel at solving structured problems without being strong general-purpose knowledge systems. This model makes that trade-off explicit: it is optimized for math, code, and instruction following, while its knowledge-heavy benchmark scores lag behind larger peers.
Key Capabilities
- Strong contest math: It scores 98.2 on MATH-500 and 88.3 on AIME 2025 in the comparisons reported. Its contest results are broadly close to Qwen3.5-9B, though individual sets vary.
- Code and instruction following: The verbose variant scores 81.6 on LiveCodeBench and 74.4 on IFBench. These results have not yet been measured on the balanced variant, so treat them as variant-specific.
- Two reasoning-length profiles: Balanced uses length-aware reinforcement learning to favor shorter traces; verbose retains longer reasoning for harder problems. The model has 8.79B parameters and supports local use, including GGUF quantizations.
Who It’s For
A good fit for people evaluating compact models on math, coding, or instruction-following tasks, especially when local inference matters. Look elsewhere for broad factual knowledge, tool use, or a general assistant: GPQA Diamond is 70.2, agentic tasks were not evaluated, and the model has no additional safety or preference tuning beyond its base model. Benchmark comparisons also combine results from different evaluation setups.
