OpenDCAI/Omni-Edu-9B-FP8

OmniEdu-9B-FP8

Open Foundation Models for Learning and Teaching

Paper · Project Page · GitHub · Dataset · Model Family · Citation

OmniEdu-9B-FP8 is the FP8-quantized release of OmniEdu-9B. The architecture, tokenizer, chat template and instruction-tuning data are unchanged: this checkpoint is the same model with its linear layers stored in FP8, which shrinks the weights from 17.53 GiB to 11.09 GiB and is intended for deployment on smaller GPUs and for higher-throughput serving.

This checkpoint accompanies OmniEdu: Open Foundation Models for Learning and Teaching.

Hao Liang · Qihan Lin · Meiyi Qiang · Linzhuang Sun · Hengyi Feng · Mingrui Chen · Sizhe Qiu · Wentao Zhang Peking University · University of the Chinese Academy of Sciences · Zhongguancun Academy

Quantization

Quantized with LLM Compressor into the compressed-tensors FP8_BLOCK scheme (library version 0.19.0), without calibration data:

Targets	every Linear layer
Weights	FP8, symmetric, 128×128 block-wise scales
Activations	FP8, dynamic, 128-element group scales
Kept in BF16	the whole vision tower (model.visual.*), the linear_attn.in_proj_a/b projections, and lm_head; every non-Linear module keeps its original dtype
Format	compressed-tensors (float-quantized), see recipe.yaml

The vision tower, the linear-attention gating projections and the output head are deliberately left unquantized, so image encoding and output logits keep full precision.

Capabilities

An educational model needs to connect what is being taught, what the learner understands, and what to do next. OmniEdu organizes supervised fine-tuning around four complementary capabilities:

Capability	What the model is trained to do
Subject competence	Solve K–12 problems across subjects and input formats
Curriculum grounding	Link questions, concepts, and solutions to curriculum standards and prerequisites
Diagnostic reasoning	Identify misconceptions, missing prerequisites, and gaps in learner understanding
Pedagogical action and scaffolding	Give targeted feedback, ask guiding questions, explain, and adapt instructional support

The training mixture contains 69,999 instruction examples and 15.96M supervised response tokens, drawn from more than 100 sources: 60,951 education-specific examples and 9,048 general-purpose examples, with 20 task-specific system instructions. The models use full-parameter supervised fine-tuning with a 32,768-token training sequence length.

Model family

Checkpoint	Scale	Weights	Backbone reported in the paper
OmniEdu-4B	4B	8.46 GiB (BF16)	Qwen3.5-4B-Base
OmniEdu-4B-FP8	4B	5.13 GiB (FP8)	Qwen3.5-4B-Base
OmniEdu-9B	9B	17.53 GiB (BF16)	Qwen3.5-9B-Base
OmniEdu-9B-FP8	9B	11.09 GiB (FP8)	Qwen3.5-9B-Base
OmniEdu-27B	27B	50.96 GiB (BF16)	Qwen3.8-27B
OmniEdu-27B-FP8	27B	28.30 GiB (FP8)	Qwen3.8-27B

This repository contains OmniEdu-9B-FP8. All published checkpoints use the Qwen3_5ForConditionalGeneration architecture. Use the multimodal model class and processor shown below, including for text-only conversations.

Evaluation

OmniEdu-9B is the BF16 original and OmniEdu-9B-FP8 is this checkpoint. Higher is better; all values are percentages except LongTutor Teaching, which uses the benchmark's original score scale. WR means win rate.

Metric ↑	OmniEdu-9B	OmniEdu-9B-FP8
K12-Bench EM	55.46%	55.62%
K12-Bench F1	73.68%	73.45%
MathFish Acc.	83.70%	81.77%
EDUMATH MaC	74.00%	73.77%
GAOKAO-Bench Full	93.66%	91.09%
EXAMS-V Overall	66.40%	63.83%
MDK12-Bench Full	50.80%	49.20%
MathTutorBench Scaffold WR	75.26%	78.11%
MathTutorBench Scaffold-hard WR	81.64%	82.42%
TutorBench Overall	48.16%	48.01%
LongTutor Evidence	66.63%	69.04%
LongTutor Teaching	2.66	2.76

Quick start

Transformers

Loading a compressed-tensors checkpoint needs the compressed-tensors package in addition to a recent Transformers release with Qwen3.5 support:

pip install -U torch torchvision transformers accelerate safetensors pillow compressed-tensors

import torch
from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration

MODEL_ID = "OpenDCAI/Omni-Edu-9B-FP8"
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-based questions, include an image item such as {"type": "image", "image": "/path/to/question.png"} alongside the text item in content. The processor prepares both modalities. See the official Qwen3.5 Transformers documentation for supported input formats.

Chat template: Training uses qwen3_5_nothink. Keep enable_thinking=False when formatting prompts to match the training setup. The 32,768-token length reported above is the training sequence length, not a measured guarantee of long-context performance.

vLLM

FP8 is the native deployment path for this format, and vLLM loads compressed-tensors checkpoints directly:

pip install -U vllm

vllm serve OpenDCAI/Omni-Edu-9B-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}'

Set --tensor-parallel-size to the number of GPUs used for the model, and reduce --max-model-len if memory is tight. See the vLLM example in the project repository for the request format.

Prompting for educational tasks

State the learner's grade, topic, prior attempt, and the kind of support you want. For example:

• Activity planning: “Design a 30-minute Grade 7 technology activity about electrical circuits using batteries, wires, and LEDs. Include learning objectives and questions to check understanding.”
• Learner diagnosis: “A student says 1/8 is greater than 1/6 because 8 is greater than 6. Identify the misconception and suggest a question that helps them reconsider.”
• Guided tutoring: “Help me solve x² − 5x + 6 = 0. Give one hint at a time and wait for my response.”

Training data

The quantization does not change the training data. Load the released instruction mixture from Hugging Face:

pip install -U datasets

from datasets import load_dataset

train = load_dataset(
    "OpenDCAI/Omni-Edu",
    split="train",
)
print(train)

The dataset covers curriculum grounding, problem solving, diagnosis, tutoring, and general instruction. Please follow the licenses and usage terms of the component datasets and source materials.

Training details

Full-parameter supervised fine-tuning uses LLaMA-Factory, BF16 precision, and the qwen3_5_nothink template, with a maximum training sequence length of 32,768 tokens. See the paper's training configuration for the full setup. The FP8 release was produced by post-training quantization of the resulting checkpoint, as described above.

Checkpoint training hyperparameters and framework versions

The following hyperparameters were used during training:

• learning_rate: 5e-06
• train_batch_size: 1
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 64
• total_eval_batch_size: 64
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 0.1
• num_epochs: 3.0

Framework versions

• Transformers 5.2.0
• Pytorch 2.10.0
• Datasets 4.0.0
• Tokenizers 0.22.2

Intended use and limitations

OmniEdu is intended for educational research, teacher-assistance tools, and learning-support prototypes. Models may produce incorrect solutions, unsupported curriculum mappings, or unsuitable instructional guidance. Educators should review generated material before using it with learners. Benchmark results do not establish improvements in classroom learning outcomes, and performance can vary by subject, curriculum, language, and image quality. Post-training quantization can shift results relative to the BF16 checkpoint; measure on your own workload before relying on the FP8 release in production.

Citation

If you use OmniEdu models or data in your research, please cite:

@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}
}

Feedback

We welcome classroom use cases, bug reports, and community contributions. Please open a discussion in this model's Community tab or an issue in the project repository.