sii-research/InnoSpark3.0-9B-260630

InnoSpark3.0-9B-260630: Education-Enhanced LLM

InnoSpark3.0-9B-260630 is a 9-billion parameter model developed by SII Research, building upon the Qwen3.5-9B base. This model is uniquely optimized for educational applications, undergoing a multi-stage training pipeline involving Supervised Fine-Tuning (SFT) with general and education-domain data, followed by Reinforcement Learning (RL) focused on reasoning, educational QA, agentic capabilities, and instruction following.

Key Capabilities & Enhancements

• Education-Specific Performance: Significantly improved scores on EduBench (+2.6 total) and Pedagogy Benchmark Multilingual (+16.08 total), demonstrating superior performance in educational contexts compared to its base model.
• Enhanced Reasoning: Achieves notable gains in reasoning benchmarks like AIME25 (+43.33) and LiveCodeBench v6 (+6.25).
• Strong Instruction Following: Shows substantial improvements in IF-Eval (+9.62) and IF-bench (+28.19).
• Agentic Capabilities: Improved performance in general agent tasks (BFCL_v4 +12.68, TAU3-bench +2.94) and coding agent tasks (Terminal-Bench 2.1 +4.34).
• General Knowledge: Maintains and slightly improves general knowledge benchmarks like MMLU-Pro (+1.49) and C-Eval (+2.83).
• Minimal Visual Performance Decrease: Despite not being specifically enhanced for vision, the model exhibits only a minor 1-2% decrease in visual performance compared to Qwen3.5-9B.

Ideal Use Cases

This model is particularly well-suited for:

• Educational QA and Tutoring: Providing explanations, scaffolded instruction, and homework support.
• Teaching Assistance: Generating lesson plans and teaching materials.
• Learning Companionship: Developing student-facing dialogue agents.
• General Assistant Tasks: Handling instruction following, reasoning, and agent-style tasks in broader contexts.