Description
OmniEdu represents an open family of foundation models meticulously developed for K–12 learning and teaching environments. This suite of models is engineered to address critical aspects of education, including curriculum grounding, K–12 problem solving, diagnostic reasoning, and pedagogical tutoring. The models are designed to be open-source, fostering innovation and collaboration within the educational technology community.
Read More
The OmniEdu family is available in multiple scales, including OmniEdu-4B for local experimentation, OmniEdu-9B for general education applications, and OmniEdu-27B, which demonstrates the strongest overall results in educational evaluations. These models are offered in both BF16 originals and FP8 releases, with FP8 versions significantly reducing download sizes by 37%–44% while maintaining performance. All three scales utilize a substantial 32,768-token training sequence length, enabling comprehensive processing of educational content.
Key capabilities of OmniEdu encompass the entire learning–teaching loop. Curriculum grounding allows the models to map questions to specific knowledge points, grade levels, prerequisite relations, and curriculum evidence. For K–12 problem solving, OmniEdu can solve authentic school-level questions across various subjects, formats, and multimodal inputs. Diagnostic reasoning helps identify misconceptions, missing prerequisites, and knowledge-state gaps from learner work, providing valuable insights into student understanding. Furthermore, pedagogical tutoring enables the models to suggest the next instructional action, such as providing a hint, asking a follow-up question, recommending a prerequisite review, or offering a detailed explanation.
The models have been rigorously trained on a vast dataset comprising 60,951 education-specific and 9,048 general-purpose examples sourced from over 100 different origins. Evaluations against 16 other models, including both open-weight and proprietary solutions, show OmniEdu-27B leading in benchmarks like K12-Bench EM/F1, MathFish accuracy, and LongTutor's teaching score, indicating strong performance in curriculum grounding and instructional support. The project emphasizes that while benchmark results are strong, educators should review model outputs before consequential use in the classroom.
OmniEdu supports educational research and the creation of teacher-assistance tools. The models are designed for deployment on GPUs/backends that support block-wise FP8, utilizing recent vLLM releases with Qwen3.5 support. Users can easily switch between different model sizes and precision levels (BF16 or FP8) by adjusting the model ID. Memory requirements vary based on context length, concurrency, and the inclusion of image inputs.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentOn-premise (via GitHub)
- PlatformsWeb
OmniEdu's Core Features
Open family of foundation models for K–12 education
Curriculum grounding: maps questions to knowledge points, grade levels, and prerequisites
K–12 problem solving: solves school-level questions across subjects and formats
Diagnostic reasoning: identifies misconceptions and knowledge gaps from learner work
Pedagogical tutoring: suggests instructional actions like hints, questions, or explanations
Multiple model scales: OmniEdu-4B, OmniEdu-9B, and OmniEdu-27B
FP8 releases: reduces download size by 37%–44% with maintained performance
BF16 originals: available for full precision experimentation
32,768-token training sequence length across all scales
Trained on 60,951 education-specific and 9,048 general-purpose examples
Strong performance in K12-Bench EM/F1, MathFish accuracy, and LongTutor teaching score
Supports educational research and teacher-assistance tool development
Deployment examples and quantization configurations provided
General capabilities remain strong after educational fine-tuning
Supports multimodal inputs for problem solving
How to use OmniEdu?
Choose Model Scale: Select OmniEdu-4B for local experimentation, OmniEdu-9B for education applications, or OmniEdu-27B for strongest overall results.
Select Precision: Opt for BF16 originals or FP8 releases, noting FP8 cuts weight downloads by 37%–44%.
Prepare Environment: Use a GPU/backend supporting block-wise FP8 and a recent vLLM release with Qwen3.5 support for deployment.
Deploy Model: Start with OmniEdu-4B-FP8, then change the model ID and adjust GPU parallelism for 9B or 27B models.
Serve BF16 Original: Remove the '-FP8' suffix from the model ID to serve the BF16 original checkpoint.
Integrate into Applications: Utilize the models for curriculum grounding, problem solving, diagnosis, or pedagogical tutoring in educational tools.
Review Outputs: Educators should review model outputs before consequential use in classroom settings.
OmniEdu's Use Cases
- Curriculum Alignment
- K–12 Problem Solving
- Learner Diagnosis
- Pedagogical Support
- Educational Research
- Teacher Assistance Tools
- Model Experimentation
- Multimodal Learning








