Open education model family OmniEdu released: 27B/9B/4B, leads proprietary models in 51 of 60 metrics
Puzzleheaded_Box2842 · reddit · 2026-09-23
OmniEdu is an open family of K-12 education foundation models (27B/9B/4B) built via full-parameter SFT on Qwen3.5/3.8 base models using LLaMA-Factory, 32k context, lr 5e-6, 3 epochs, DeepSpeed ZeRO-3. Training data: 69,999 examples (61k education-specific, curated from 1.34M candidates through a six-stage pipeline), 15.96M supervised tokens, organized around subject competence, curriculum grounding, diagnostic reasoning, and pedagogical scaffolding with 20 task-specific system prompts. It leads in 51 of 60 metric-level comparisons against five proprietary models; OmniEdu-27B lifts K12-Bench EM from 52.11% to 63.12% and MathTutorBench Scaffold from 57.16% to 78.74%; the 4B variant jumps scaffold win rate from 20.42% to 75.79%. Authors caution benchmarks don't directly evidence student learning gains.
Related event: Peking University Open-Sources OmniEdu Education Models(2 posts)→
More from Models
- Opus 5.5 One-Shots a Full Prince of Persia Level With Graphics, NPCs and Music — iannuttall · 2026-09-23
- Opus 5.5 Cut Prices 40% and Within a Day It Was Porting C to Rust: 8 Use Cases — alex_verem · 2026-09-23
- Beff Jezos jokes Opus 5.5 is 'post-slop', freeing readers from AI sludge prose — beffjezos · 2026-09-23
- Opus 5.5 One-Shots a Full Prince of Persia Level with NPCs, Sound and Music — iannuttall · 2026-09-23
- GPT-6 Sol Underperforms GPT-5.6 Max on DeepSWE, 68.8% vs 72.7% — banaxi-tech · 2026-09-23
- Model release fatigue: devs say gains are marginal, but some argue new models clearly outpace old ones — Rasmic · 2026-09-23