OmniEdu-4B
Our smallest checkpoint for local experimentation.
39% smaller weight download with FP869,999instruction examples
15.96Msupervised response tokens
3open model scales
20task-specific system instructions
01 / OPEN MODEL FAMILY
Choose your scale.
Start building.
BF16 originals and FP8 releases at every scale. FP8 cuts weight downloads by 37%–44%. All three scales use a 32,768-token training sequence length.
October 2026 · FP8 release Six checkpoints, with quantized results and deployment examples. Compare BF16 and FP8 ↓
OmniEdu-9B
The middle scale for education applications.
37% smaller weight download with FP8OmniEdu-27B
Our strongest overall results across the education evaluations.
44% smaller weight download with FP8Sizes are summed safetensors downloads, checked October 7, 2026. Runtime GPU memory also includes caches, activations, and serving overhead. FP8 preserves the vision tower, linear-attention gating projections, and output head in their original precision.
Trained on 60,951 education-specific and 9,048 general-purpose examples from more than 100 sources. Explore the dataset ↗
02 / CAPABILITIES
One family for the full learning–teaching loop.
01
Curriculum grounding
Map a question to knowledge points, grade levels, prerequisite relations, and curriculum evidence.
02
K–12 problem solving
Solve authentic school-level questions across subjects, formats, and multimodal inputs.
03
Diagnostic reasoning
Identify misconceptions, missing prerequisites, and knowledge-state gaps from learner work.
04
Pedagogical tutoring
Choose the next instructional action: a hint, question, prerequisite review, or explanation.
03 / EVALUATION
Leading curriculum grounding.
Stronger instructional support.
Among the 16 models evaluated in the paper, OmniEdu-27B leads K12-Bench EM/F1, MathFish accuracy, and LongTutor's teaching score. Explore the full comparison with open-weight and proprietary models below.
K12-Bench · OmniEdu-27B63.12% EM76.69% F1 · best on both metrics
MathFish · OmniEdu-27B85.89%Best accuracy among evaluated models
LongTutor · OmniEdu-27B3.02Best teaching score · original scale
All models. All three education evaluations.
Higher is better. Per-column top three: 1st · 2nd · 3rd. Shaded rows identify OmniEdu.
Curriculum grounding16 models · Table 1
| Model | Size | K12-Bench EM | K12-Bench F1 | MathFish Acc. | EDUMATH MaC |
|---|---|---|---|---|---|
| Open-weight Models | |||||
| Qwen3.5-4B-Base | 4B | 42.72% | 69.20% | 80.33% | 47.60% |
| OmniEdu-4B (ours) | 4B | 54.25% | 71.75% | 83.19% | 68.40% |
| Qwen3.5-9B-Base | 9B | 48.52% | 71.99% | 79.66% | 58.60% |
| OmniEdu-9B (ours) | 9B | 55.46% | 73.68% | 83.70% | 74.00% |
| Qwen3.8-27B | 27B | 52.11% | 73.48% | 83.54% | 70.60% |
| OmniEdu-27B (ours) | 27B | 63.12% | 76.69% | 85.89% | 86.95% |
| Confucius3-Math | 14B | 5.23% | 19.23% | 0.00% | 3.00% |
| MuduoLLM | 14B | 48.24% | 70.46% | 82.85% | 57.20% |
| EduChat-SFT-Qwen2.5-7B | 7B | 45.20% | 67.58% | 78.23% | 28.00% |
| EduChat-R1-Qwen3-8B | 8B | 25.25% | 49.67% | 19.49% | 47.00% |
| EduChat-R1-Qwen3-32B | 32B | 49.42% | 71.21% | 83.71% | 61.20% |
| Proprietary Models | |||||
| GPT-5.4 | – | 43.23% | 67.80% | 80.65% | 80.50% |
| GPT-5.6-Sol | – | 48.63% | 71.03% | 83.68% | 84.92% |
| Claude-Opus-5 | – | 48.48% | 71.34% | 83.52% | 75.00% |
| GLM-5.3 | – | 43.38% | 57.66% | 85.23% | 62.78% |
| Kimi-K3 | – | 55.03% | 74.71% | 85.40% | 90.00% |
K–12 problem solving16 models · Table 2
| Model | Size | GAOKAO-Bench Full | EXAMS-V Overall | MDK12-Bench Full |
|---|---|---|---|---|
| Open-weight Models | ||||
| Qwen3.5-4B-Base | 4B | 88.98% | 44.69% | 35.43% |
| OmniEdu-4B (ours) | 4B | 90.46% | 57.62% | 46.36% |
| Qwen3.5-9B-Base | 9B | 92.94% | 63.00% | 44.50% |
| OmniEdu-9B (ours) | 9B | 93.66% | 66.40% | 50.80% |
| Qwen3.8-27B | 27B | 91.55% | 68.65% | 46.04% |
| OmniEdu-27B (ours) | 27B | 94.87% | 69.52% | 57.76% |
| Confucius3-Math | 14B | 84.51% | 21.95% | 39.13% |
| MuduoLLM | 14B | 89.66% | 21.95% | 45.44% |
| EduChat-SFT-Qwen2.5-7B | 7B | 70.75% | 21.95% | 32.50% |
| EduChat-R1-Qwen3-8B | 8B | 76.08% | 0.69% | 41.29% |
| EduChat-R1-Qwen3-32B | 32B | 83.86% | 21.95% | 41.84% |
| Proprietary Models | ||||
| GPT-5.4 | – | 93.44% | 35.66% | 54.77% |
| GPT-5.6-Sol | – | 95.96% | 24.82% | 54.67% |
| Claude-Opus-5 | – | 97.22% | 64.34% | 57.46% |
| GLM-5.3 | – | 89.76% | 22.29% | 55.33% |
| Kimi-K3 | – | 94.86% | 87.29% | 63.60% |
Pedagogical tutoring16 models · Table 3
| Model | Size | MathTutor-S | MathTutor-SH | TutorBench | LongTutor-E | LongTutor-T |
|---|---|---|---|---|---|---|
| Open-weight Models | ||||||
| Qwen3.5-4B-Base | 4B | 20.42% | 18.36% | 45.52% | 25.67% | 1.60 |
| OmniEdu-4B (ours) | 4B | 75.79% | 84.77% | 46.67% | 65.88% | 2.29 |
| Qwen3.5-9B-Base | 9B | 14.00% | 13.67% | 45.38% | 5.81% | 1.48 |
| OmniEdu-9B (ours) | 9B | 75.26% | 81.64% | 48.16% | 66.63% | 2.66 |
| Qwen3.8-27B | 27B | 57.16% | 55.86% | 58.58% | 36.80% | 2.74 |
| OmniEdu-27B (ours) | 27B | 78.74% | 83.59% | 59.42% | 78.20% | 3.02 |
| Confucius3-Math | 14B | 27.68% | 20.70% | 34.72% | 23.07% | 1.36 |
| MuduoLLM | 14B | 30.42% | 23.44% | 35.98% | 24.31% | 1.28 |
| EduChat-SFT-Qwen2.5-7B | 7B | 18.21% | 22.66% | 21.48% | 26.55% | 1.21 |
| EduChat-R1-Qwen3-8B | 8B | 11.58% | 7.42% | 23.71% | 33.75% | 1.14 |
| EduChat-R1-Qwen3-32B | 32B | 22.21% | 16.80% | 25.92% | 52.03% | 1.56 |
| Proprietary Models | ||||||
| GPT-5.4 | – | 6.32% | 1.96% | 46.34% | 75.50% | 1.62 |
| GPT-5.6-Sol | – | 10.53% | 17.65% | 50.97% | 77.50% | 1.89 |
| Claude-Opus-5 | – | 87.89% | 84.31% | 54.01% | 84.17% | 2.72 |
| GLM-5.3 | – | 77.89% | 76.92% | 38.62% | 77.42% | 2.66 |
| Kimi-K3 | – | 20.00% | 15.38% | 63.65% | 74.73% | 2.17 |
OmniEdu entries use the original BF16 checkpoints. Scores reproduce arXiv v1 (Tables 1–3); higher is better. Values are percentages except LongTutor-T, which uses the original teaching-score scale. MathTutor-S / SH denote Scaffold / Scaffold-hard win rates; LongTutor-E / T denote Evidence / Teaching averages. The five education-specific baselines and GLM-5.3 receive text-only inputs on image-dependent examples, as described in the paper. Rankings apply to the evaluation protocol reported in the paper. Evaluation protocol ↗
Smaller weights. Benchmark by benchmark.
Compare the BF16 originals with the published FP8 results at every scale. These model-card results are separate from the paper's 16-model ranking; scores can increase or decrease after quantization.
FP8 · Curriculum grounding3 scales · BF16 vs. FP8
| Model | K12-Bench EM | K12-Bench F1 | MathFish Acc. | EDUMATH MaC |
|---|---|---|---|---|
| OmniEdu-4B BF16 | 54.25% | 71.75% | 83.19% | 68.40% |
| OmniEdu-4B FP8 | 53.69% | 70.94% | 82.26% | 67.64% |
| OmniEdu-9B BF16 | 55.46% | 73.68% | 83.70% | 74.00% |
| OmniEdu-9B FP8 | 55.62% | 73.45% | 81.77% | 73.77% |
| OmniEdu-27B BF16 | 63.12% | 76.69% | 85.89% | 86.95% |
| OmniEdu-27B FP8 | 63.16% | 76.59% | 85.13% | 86.84% |
FP8 · K–12 problem solving3 scales · BF16 vs. FP8
| Model | GAOKAO-Bench Full | EXAMS-V Overall | MDK12-Bench Full |
|---|---|---|---|
| OmniEdu-4B BF16 | 90.46% | 57.62% | 46.36% |
| OmniEdu-4B FP8 | 89.47% | 57.14% | 47.30% |
| OmniEdu-9B BF16 | 93.66% | 66.40% | 50.80% |
| OmniEdu-9B FP8 | 91.09% | 63.83% | 49.20% |
| OmniEdu-27B BF16 | 94.87% | 69.52% | 57.76% |
| OmniEdu-27B FP8 | 95.72% | 69.15% | 58.81% |
FP8 · Pedagogical tutoring3 scales · BF16 vs. FP8
| Model | MathTutor-S | MathTutor-SH | TutorBench | LongTutor-E | LongTutor-T |
|---|---|---|---|---|---|
| OmniEdu-4B BF16 | 75.79% | 84.77% | 46.67% | 65.88% | 2.29 |
| OmniEdu-4B FP8 | 76.63% | 83.20% | 46.15% | 66.11% | 2.30 |
| OmniEdu-9B BF16 | 75.26% | 81.64% | 48.16% | 66.63% | 2.66 |
| OmniEdu-9B FP8 | 78.11% | 82.42% | 48.01% | 69.04% | 2.76 |
| OmniEdu-27B BF16 | 78.74% | 83.59% | 59.42% | 78.20% | 3.02 |
| OmniEdu-27B FP8 | 76.11% | 80.08% | 59.34% | 78.95% | 3.05 |
Higher is better. Values are percentages except LongTutor-T (Teaching); LongTutor-E is Evidence. MathTutor-S / SH denote Scaffold / Scaffold-hard win rates. Shaded rows identify FP8 checkpoints. Source: 4B model card ↗ · 9B model card ↗ · 27B model card ↗.
General capabilities remain strong.
The original BF16 checkpoints improve at all three sizes on the selected metrics below after educational fine-tuning. Individual submetrics can be flat or decline; the appendix reports the full breakdown.
| Model | IFEval Prompt Strict | GPQA Diamond | MMMU-Pro Overall |
|---|---|---|---|
| Qwen3.5-4B-Base | 64.88% | 58.08% | 50.46% |
| OmniEdu-4B (ours) | 65.06% | 60.61% | 52.60% |
| Qwen3.5-9B-Base | 69.69% | 62.12% | 58.38% |
| OmniEdu-9B (ours) | 73.01% | 63.64% | 60.75% |
| Qwen3.8-27B | 80.59% | 74.24% | 64.97% |
| OmniEdu-27B (ours) | 82.07% | 77.78% | 67.98% |
IFEval: prompt-level strict accuracy. GPQA: Diamond accuracy. MMMU-Pro: overall accuracy. Values are percentages; sources: Tables 13–15 of arXiv v1.
04 / DEPLOYMENT
From checkpoint
to your application.
Start with OmniEdu-4B-FP8 on a GPU/backend supporting block-wise FP8, using a recent vLLM release with Qwen3.5 support. The checkpoint carries its quantization configuration; thinking is disabled to match training.
For 9B or 27B, change the model ID and adjust GPU parallelism. To serve the BF16 original, remove the -FP8 suffix. Memory needs depend on context length, concurrency, and image inputs.
pip install -U vllm
vllm serve lhpku20010120/Omni-Edu-4B-FP8 \
--served-model-name omniedu \
--dtype bfloat16 \
--tensor-parallel-size 1 \
--max-model-len 32768 \
--reasoning-parser qwen3 \
--default-chat-template-kwargs '{"enable_thinking": false}'05 / THE PAPER
OmniEdu: Open Foundation Models
for Learning and Teaching
Hao Liang · Qihan Lin · Meiyi Qiang · Linzhuang Sun · Hengyi Feng · Mingrui Chen · Sizhe Qiu · Wentao Zhang
arXiv:2609.23088 · September 19, 2026
Cite OmniEdu · BibTeX
@misc{liang2026omniedu,
title = {OmniEdu: Open Foundation Models for Learning and Teaching},
author = {Hao Liang and Qihan Lin and Meiyi Qiang and Linzhuang Sun and Hengyi Feng and Mingrui Chen and Sizhe Qiu and Wentao Zhang},
year = {2026},
eprint = {2609.23088},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2609.23088}
}OmniEdu supports educational research and teacher-assistance tools. Benchmark results do not establish classroom learning gains; educators should review model outputs before consequential use.
“
An educational model must coordinate what to teach, where the learner is, and how to respond.
— OmniEdu design principle