Releases·Model Release
OmniEdu: Open Foundation Models for Learning and Teaching — 27B / 9B / 4B
The OmniEdu team released an open family of K-12 education foundation models built by full-parameter supervised fine-tuning of Qwen3.5/Qwen3.8 base models. The models target four capabilities: subject competence, curriculum grounding, diagnostic reasoning, and pedagogical scaffolding. OmniEdu-27B leads in 51 of 60 metric-level comparisons against five proprietary models, with large gains in tutoring benchmarks over its base model.
1 engagement1 sourceWed, Sep 23, 2026, 07:46 AM
AI Overview
- Team releases OmniEdu open K–12 education models in 27B/9B/4B sizes, fine-tuned from Qwen3.5/3.8 bases on 69,999 examples (60,951 education-specific).
- Training uses LLaMA-Factory, 32,768 sequence length, 5e-6 learning rate, three epochs, BF16, DeepSpeed ZeRO-3 across all three variants.
- Data pipeline filters 1.34M education candidates down to 60,951 examples across four capabilities: subject competence, curriculum grounding, diagnostic reasoning, pedagogical scaffolding.
- OmniEdu-27B leads in 51 of 60 metric-level comparisons against five proprietary models; K12-Bench EM rises from 52.11% to 63.12% over its base model.
- OmniEdu-4B lifts MathTutorBench Scaffold win rate from 20.42% to 75.79%; MMMU-Pro accuracy on the 27B variant improves from 64.97% to 67.98%.
Related
Releases·Model Release
Qwen3.8-27B now runs on Cerebras with rapid inference
Qwen3.8-27B is now available on Cerebras' inference platform, announced via an X post promoting its fast, open-weight serving. This gives developers access to the open Qwen model with Cerebras' high-speed inference.
Qwen3.8-27B is now available on Cerebras' inference platform, announced via an X post promoting its fast, open-weight serving. This gives developers access to the open Qwen model with Cerebras' high-speed inference.
1.0k engagement1 sourcexSat, Sep 12, 2026, 07:35 AM
Products·building
Local LLM agent system that automates directory submission form filling
A self-hosted automation pipeline where an uncensored Qwen3.8-27B orchestrator spawns isolated worker processes, each driving a headless Chrome instance to fill out web directory submission forms. Workers handle varying form fields and solve captchas via a vision model. It solves the tedious manual task of submitting a site to dozens of directories.
A self-hosted automation pipeline where an uncensored Qwen3.8-27B orchestrator spawns isolated worker processes, each driving a headless Chrome instance to fill out web directory submission forms. Workers handle varying form fields and solve captchas via a vision model. It solves the tedious manual task of submitting a site to dozens of directories.
7 engagement1 sourceredditFri, Oct 16, 2026, 08:00 PM
Developments·Paper
LittleLearner 5B LLM trained on K-5 curriculum fails to generalize
Researchers released LITTLECURRICULUM, an 88B-token corpus of U.S. elementary school material, and trained LITTLELEARNER (0.6B, 1.3B, 5B) from scratch. The 5B model shows sufficient language competence but fails to generalize beyond its training distribution, unlike Talkie, a 13B model trained on pre-1931 text that could write Python code.
Researchers released LITTLECURRICULUM, an 88B-token corpus of U.S. elementary school material, and trained LITTLELEARNER (0.6B, 1.3B, 5B) from scratch. The 5B model shows sufficient language competence but fails to generalize beyond its training distribution, unlike Talkie, a 13B model trained on pre-1931 text that could write Python code.
120480 engagement3 sourcesarxiv, hackernews, redditThu, Aug 13, 2026, 05:56 PM
Developments·Benchmark
ELBench: A Multi-Dimensional Benchmark for Education-Facing Large Language Models
Researchers introduced ELBench, the first benchmark to evaluate education-facing LLMs across four integrated requirements: general capability, safety, instructional usefulness, and pedagogical alignment. Existing benchmarks assess these dimensions in isolation, so ELBench provides a unified profile for suitability in tutoring, teaching assistance, and content generation.
Researchers introduced ELBench, the first benchmark to evaluate education-facing LLMs across four integrated requirements: general capability, safety, instructional usefulness, and pedagogical alignment. Existing benchmarks assess these dimensions in isolation, so ELBench provides a unified profile for suitability in tutoring, teaching assistance, and content generation.
0 engagement2 sourcesarxivMon, Aug 10, 2026, 12:46 PM
Developments·Paper
FATE: A Specialized 8B-Parameter Model for Evaluating AI Tutors
Researchers introduce FATE (FLC AI Tutor Evaluator), an 8B-parameter language model fine-tuned via knowledge distillation to evaluate AI tutors across four pedagogical tracks: Mistake Identification, Mistake Location, Guidance, and Actionability. The model aligns with the BEA 2025 Shared Task and aims to address the lack of reliable automated evaluation methods for LLM-based tutors in education.
Researchers introduce FATE (FLC AI Tutor Evaluator), an 8B-parameter language model fine-tuned via knowledge distillation to evaluate AI tutors across four pedagogical tracks: Mistake Identification, Mistake Location, Guidance, and Actionability. The model aligns with the BEA 2025 Shared Task and aims to address the lack of reliable automated evaluation methods for LLM-based tutors in education.
0 engagement1 sourcearxivSun, Jul 12, 2026, 08:28 AM