Model Details
Table with columns: Item, Description| Item | Description |
|---|
| Model name | InnoSpark3.0-397B-260712 |
| Training source | NEX 397B model family |
| Reference baselines | Qwen3.5-397B-A17B, Nex-N2-Pro, SOTA models from the evaluation sheet |
| Parameters | 397B |
| Training pipeline | SFT + multi-stage RL |
| SFT data | General data + education-domain data |
| RL focus | Reasoning, educational QA, agent scenarios, instruction following |
| Primary use cases | Education QA, tutoring, teaching support, educational agents, general assistant tasks |
Training
The post-training pipeline contains two major stages:
- Supervised Fine-Tuning (SFT): uses a mixture of general instruction data and education-domain data to strengthen the model's ability to answer pedagogical questions, explain concepts, and follow classroom-oriented instructions.
- Multi-stage Reinforcement Learning (RL): improves reasoning, education-specific QA, agentic task solving, and instruction-following robustness through staged optimization.
Evaluation
All scores below are normalized to a 100-point scale. For EduBench, the original scores in the evaluation sheet are on a 10-point scale and are multiplied by 10 here. Bold values indicate the best score in each row. Missing results are shown as -.
General Benchmarks
Table with columns: Type, Capability, Benchmark, Qwen3.5-397B-A17B, Nex-N2-Pro, InnoSpark3.0-397B, Nex-N2-Pro-math-rl, GLM-5.2, DeepSeekV4-Pro, Kimi-2.6, GPT-5.5, Gemini-3.1-Pro, Claude-opus-4.8| Type | Capability | Benchmark | Qwen3.5-397B-A17B | Nex-N2-Pro | InnoSpark3.0-397B | Nex-N2-Pro-math-rl | GLM-5.2 | DeepSeekV4-Pro | Kimi-2.6 | GPT-5.5 | Gemini-3.1-Pro | Claude-opus-4.8 |
|---|
| Language | Knowledge | MMLU-Pro | 87.8 | 87.76 | 87.83 |
Education Benchmarks
The education evaluation covers EduBench and Pedagogy-oriented evaluation settings. The table below reports the detailed EduBench and Pedagogy Benchmark Multilingual metrics provided in the evaluation sheet.
Table with columns: Benchmark, Metric, Qwen3.5-397B-A17B, Nex-N2-Pro, InnoSpark3.0-397B, Nex-N2-Pro-math-rl, GLM-5.2, DeepSeekV4-Pro, Kimi-2.6, GPT-5.5, Gemini-3.1-Pro, Claude-opus-4.8| Benchmark | Metric | Qwen3.5-397B-A17B | Nex-N2-Pro | InnoSpark3.0-397B | Nex-N2-Pro-math-rl | GLM-5.2 | DeepSeekV4-Pro | Kimi-2.6 | GPT-5.5 | Gemini-3.1-Pro | Claude-opus-4.8 |
|---|
| EduBench | Total | 91.2 | 90.6 | 96.3 | - | 93.5 |
Usage
import torch
from transformers import AutoProcessor, AutoModelForMultimodalLM
model_id = "sii-research/InnoSpark3.0-397B-260712"
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForMultimodalLM.from_pretrained(
model_id,
torch_dtype="auto",
device_map="auto",
trust_remote_code=True,
)
messages = [
{
"role": "user",
"content": [
{
"type": "text",
"text": "请用适合初中生的方式讲解一元二次方程的求根公式，并给出一个例题。",
}
],
}
]
inputs = processor.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt",
).to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024)
response = processor.decode(
outputs[0][inputs["input_ids"].shape[-1]:],
skip_special_tokens=True,
)
print(response)
vLLM
vllm serve sii-research/InnoSpark3.0-397B-260712 --trust-remote-code
Intended Use
InnoSpark3.0-397B-260712 is intended for research and application development in education-focused AI scenarios, including:
- Concept explanation and step-by-step tutoring
- Educational QA and homework support
- Lesson planning and teaching material generation
- Student-facing dialogue agents
- Teacher-facing assistant workflows
- General instruction following, reasoning, and agent-style tasks
Limitations
Like other large language models, InnoSpark3.0-397B-260712 may generate inaccurate, incomplete, or biased content. Outputs in educational settings should be reviewed by qualified educators when used for high-stakes learning, assessment, or student guidance. The model should not be used as the sole source for factual verification, grading decisions, psychological counseling, medical advice, legal advice, or other safety-critical decisions.
Evaluation results may vary with prompt format, decoding parameters, evaluation implementation, and data version. Users should conduct additional evaluations before deploying the model in production or classroom environments.
Main Contributions
Table with columns: Name, Responsibility, Personal link| Name | Responsibility | Personal link |
|---|
| Wentao Liu (刘文涛) | Training pipeline; SFT general and education data processing; education RL training | Google Scholar |
| Siyu Song (宋思宇) | RL training environment infrastructure; general RL training; education RL training | Google Scholar |
| Ye Lu (卢烨) | SFT training; RL for instruction-following capability | GitHub, Google Scholar |
Citation
@misc{innospark3_397b_260712,
title = {InnoSpark3.0-397B-260712},
author = {SII Research},
year = {2026},
howpublished = {\url{https://huggingface.co/sii-research/InnoSpark3.0-397B-260712}}
}
Please also follow the citation and license requirements of the related base and reference models.