Education-domain LLM series developed jointly by SII and East China Normal University; 3.0 is the first multimodal generation, an "education-enhanced 9B model developed from Qwen3.5-9B" (image-text-to-text, 262K context) trained with SFT plus multi-stage RL. Apache-2.0. Card-reported gains over the base include EduBench 87.4 (vs 84.8) and Pedagogy Benchmark Multilingual 84.8 (vs 68.7); the larger reasoning deltas (AIME25/26 93.3 vs 50) are self-reported and likely compare thinking-mode output against the raw base — no technical report yet.

Model Details

Context window 262,144
License Apache 2.0
Base model qwen3.5

Variants

Name Parameters Notes
InnoSpark3.0-9B (260630) 9B
InnoSpark3.0-35B (260709) 35B MoE scale-up released July 9, 2026; card reports mixed wins vs Qwen3.6-35B-A3B (MMMU-Pro 81.0, IF-Eval 92.2); base model undeclared on the card.
educationdomain-specificopen-weight