Skip to content
OmniEdu: Open-Source K-12 Educational Foundation Model with Diagnostic Reasoning

OmniEdu: Open-Source K-12 Educational Foundation Model with Diagnostic Reasoning

Peking University researchers release OmniEdu, an open-source K-12 educational foundation model family (4B, 9B, 27B) built on Qwen3.5/3.8, featuring curriculum alignment, diagnostic reasoning, and pedagogical scaffolding, with full paper, weights, training data, code, and evaluation toolkit, outperforming base models on education benchmarks while retaining general capabilities.

Code Farming
Code Farming
Code Farming

Overview

OmniEdu is a newly open-sourced K-12 educational foundation model family developed by researchers from Peking University, the University of Chinese Academy of Sciences, and Zhongguancun Academy. The project releases paper (arXiv:2609.23088), model weights, training data, training code, and an evaluation toolkit under a single GitHub repository: https://github.com/haolpku/Omni-Edu.

The model family comes in three sizes — 4B, 9B, and 27B parameters — each based on Qwen3.5 or Qwen3.8 base models (OmniEdu-4B on Qwen3.5-4B-Base, OmniEdu-9B on Qwen3.5-9B-Base, OmniEdu-27B on Qwen3.8-27B). All checkpoints share the same architecture class Qwen3_5ForConditionalGeneration and are trained with the qwen3_5_nothink chat template; inference must set enable_thinking=False.

Why General Models Fail at Teaching

The article illustrates the gap with a concrete example: a Grade 7 student says "summer is warmer because Earth is closer to the Sun." A general LLM would simply correct the answer, but an educational model should first identify the misconception ("distance determines temperature"), then pose a guiding question before explaining. The author argues that solving a problem and teaching it are different tasks. General models excel at "giving answers" but lack three educational capabilities: aligning questions to curriculum standards and prerequisite knowledge, diagnosing whether a student's error stems from a misconception or missing prior knowledge, and providing targeted feedback instead of restating solutions.

Four Training Objectives

OmniEdu organizes supervised fine-tuning around four intertwined capabilities (not independent modules):

Subject problem-solving ability — cross-disciplinary, multi-format K-12 problem solving.

Curriculum alignment ability — mapping questions to standards and prerequisite concepts.

Diagnostic reasoning ability — distinguishing conceptual errors from knowledge gaps.

Teaching action and scaffolding ability — generating guiding questions, explanations, and dynamically adjusting support level.

The README emphasizes these are training targets realized via data mixing in a single model, not separate switches.

Training Recipe and Model Variants

The training mixture comprises 69,999 instruction samples (15.96M supervised tokens) from over 100 sources: 60,951 education-specific samples and 9,048 general samples, each with task-relevant system prompts. Full-parameter supervised fine-tuning uses a sequence length of 32,768.

In October, FP8 quantized checkpoints were added using LLM Compressor's FP8_BLOCK scheme: 128×128 block-wise weight quantization with dynamic FP8 activations. The vision tower, linear attention in_proj_a/b projections, and lm_head are kept in higher precision. FP8 weights are 39%, 37%, and 44% smaller for 4B, 9B, and 27B respectively. However, the README warns that smaller files do not guarantee reduced inference memory; Transformers may dequantize weights back to BF16 at runtime. The authors recommend using vLLM for FP8 serving rather than relying on file size alone.

Benchmark Results

The paper's 16-model comparison table shows OmniEdu-27B ranking first on K12-Bench EM (63.12% vs. base 52.11%), K12-Bench F1 (76.69% vs. 73.48%), MathFish accuracy (85.89%), and LongTutor teaching score (3.02 vs. 2.74). On EDUMATH MaC it scores 86.95% (second among all evaluated models), and on MathTutorBench Scaffold win-rate 78.74% (also second). GAOKAO-Bench reaches 94.87%.

Crucially, general capabilities are preserved or slightly improved: IFEval strict accuracy 82.07%, GPQA Diamond 77.78%, MMMU-Pro 67.98% — all above the base model.

The article honestly notes limitations: Kimi-K3 leads on EDUMATH MaC (90.00%) and EXAMS-V (87.29%); Claude-Opus-5 leads on MathTutorBench Scaffold (87.89%). OmniEdu's advantage lies specifically in K-12 curriculum alignment and long-horizon tutoring scores, not across every benchmark.

Deployment Options and Tooling

Developers can choose model size (4B/9B/27B) and precision (BF16 or FP8). The README clarifies that reported sizes are Hugging Face safetensors totals (verified 2026-10-07); runtime GPU memory must also account for KV cache, activations, vision inputs, and serving overhead.

A Python package omniedu (on PyPI) provides a unified client for OmniEdu and any OpenAI-compatible endpoint, plus an evaluation suite covering 12 education and general benchmarks with open-response scoring and Markdown comparison tables. Benchmark datasets and upstream code often need separate acquisition; pip install omniedu does not bundle them.

Two official deployment paths:

Transformers minimal example (quick research reproduction):

pip install -U torch torchvision transformers accelerate safetensors pillow
import torch
from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration
MODEL_ID = "lhpku20010120/Omni-Edu-4B"  # or Omni-Edu-9B / Omni-Edu-27B
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = Qwen3_5ForConditionalGeneration.from_pretrained(
    MODEL_ID,
    dtype=torch.bfloat16,
    device_map="auto",
).eval()
messages = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "A Grade 7 student says summer is warmer because Earth is closer to the Sun. Identify the misconception and give a guiding question before explaining."}
        ]
    }
]
inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    enable_thinking=False,
).to(model.device)
with torch.inference_mode():
    outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)
answer = processor.batch_decode(
    outputs[:, inputs["input_ids"].shape[-1]:],
    skip_special_tokens=True,
)[0]
print(answer)
# For image questions, add an "image" item in content; processor handles both.

vLLM OpenAI-compatible service (production integration):

pip install -U vllm
vllm serve lhpku20010120/Omni-Edu-4B \
    --served-model-name omniedu-4b \
    --dtype bfloat16 \
    --tensor-parallel-size 1 \
    --max-model-len 32768 \
    --reasoning-parser qwen3 \
    --default-chat-template-kwargs '{"enable_thinking": false}'
# For 9B/27B, change model ID and --served-model-name accordingly.
# For FP8, use the -FP8 suffixed model ID.
# Reduce --max-model-len if GPU memory is insufficient.

Evaluation toolkit installation:

pip install omniedu
pip install "omniedu[eval]"
omniedu eval --list

Training code and datasets are in the repository. A dry-run for the 9B configuration: bash training/run_training.sh 9b --dry-run Datasets are hosted on Hugging Face ( lhpku20010120/Omni-Edu) covering curriculum alignment, problem solving, diagnosis, tutoring, and general instructions. Users must respect each constituent dataset's license. Code is Apache-2.0; weights and third-party datasets/benchmarks retain their original licenses.

Boundaries and Known Limitations

Platform, not product: OmniEdu is a model and Python library, not a ready-to-use application. It requires compatible GPUs and sufficient VRAM; FP8 needs backend support for block-wise FP8 quantization. No GUI screenshots are provided because there is no product interface.

Benchmark scores ≠ classroom effectiveness: High scores on K12-Bench, MathFish, LongTutor, etc., do not prove learning gains. The model may produce incorrect answers or inappropriate guidance; teachers must review outputs before high-stakes use.

Evaluation protocol caveats: The 16-model comparison includes five education-specific baselines plus GLM-5.3. On image-dependent samples, those baselines received only text input (per paper), so rankings apply only under that specific protocol, not a fully fair multimodal contest.

FP8 metrics are point estimates: FP8 scores come from model cards, not the paper's 16-model ranking. Quantization causes some metrics to rise and others to fall; the authors do not claim statistical equivalence between BF16 and FP8.

Positioning difference: OmniEdu targets the "open-source, self-deployable, paper+data+code fully public" route. Closed-source API-only education products are a different category; on absolute problem-solving accuracy, some closed models still lead on certain metrics.

Licensing and Access

Code: Apache-2.0

Weights: retain respective licenses

Third-party datasets and benchmarks: retain their own licenses

GitHub: https://github.com/haolpku/Omni-Edu Hugging Face models: lhpku20010120/Omni-Edu-4B, lhpku20010120/Omni-Edu-9B, lhpku20010120/Omni-Edu-27B (and FP8 variants)

Hugging Face dataset: lhpku20010120/Omni-Edu arXiv paper: 2609.23088

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

model deploymentopen-source LLMbenchmark evaluationFP8 quantizationQwen3.5educational AIK-12 educationOmniEdu
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.