InnoSpark3.0-397B-260712 is an education-enhanced 397B model in the InnoSpark3.0 series. It is trained from the NEX 397B model family and evaluated against Qwen3.5-397B-A17B, Nex-N2-Pro, and additional SOTA models provided in the evaluation sheet.

InnoSpark3.0-397B-260712 is designed for educational QA, teaching assistance, learning companionship, and classroom or homework scenarios, including explanation generation, scaffolded instruction, guided reasoning, and pedagogical strategy suggestions. We use both general and education-domain data during SFT, and further strengthen reasoning, educational QA, agentic capabilities, and instruction following through a multi-stage RL pipeline. We did not specifically optimize visual capabilities; vision-language metrics are reported for transparency.

Model Details

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

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	-	84.3	73.5	86.87	-	91	-
Language	Knowledge	C-Eval	93	93.24	93.68	-	92.27	93.1	95.17	-	-	-
Language	Knowledge	SimpleQA-Verified	53.1	59.5	60.4	-	34.1	55.2	41.1	-	75.6	-
Language	Instruction Following	IF-Eval	92.6	94	90.39	-	90.39	91.9	94.5	-	-	-
Language	Instruction Following	IF-bench	76.5	64.97	70.67	-	73.3	68.03	64.63	-	-	-
Language	STEM & Reasoning	GPQA Diamond	85.86	90.7	90.4	-	91.2	90.1	90.5	93.6	94.3	-
Language	STEM & Reasoning	LiveCodeBench v6	83.6	67.39	66.54	-	54.12	93.5	89.6	-	91.7	-
Language	STEM & Reasoning	AIME25	93.33	96.67	93.33	90	80	90	93.33	-	96.67	-
Language	STEM & Reasoning	AIME26	91.3	90	96.67	93.33	99.2	96.67	96.4	98.3	100	95.7
Language	Coding Agent	SWE-bench Verified-Agentic	76.4	80.8	78.4	-	82	80.6	80.2	82.9	80.6	-
Language	Coding Agent	Terminal-Bench 2.1	52.5	75.3	57.3	-	81	67.9	66.7	83.4	68.5	74.6
Language	General Agent	BFCL_v4	72.9	64.14	65.16	-	76.66	72.95	67.82	-	-	-
Language	General Agent	TAU3-bench	68.3	71.1	60.43	-	47.73	56.63	64.7	-	50.67	-
Vision-Language	STEM & Puzzle	MMMU-Pro	85.9	85.72	86.24	-	-	-	82.6	83.2	83	-
Vision-Language	General VQA	MMBenchEN-DEV-v1.1	92.93	93.3	93.44	-	-	-	93.44	-	-	-
Vision-Language	Document Understanding	OCRBench	91	87.5	88	-	-	-	90.4	-	-	-

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.

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	91.9	93.2	92.7	94.5	90.6
EduBench	Instruction Following & Task Completion	93.7	93.7	98.2	-	96.1	96.8	90.5	98.1	97.8	86.5
EduBench	Role & Tone Consistency	95.1	94.2	99.1	-	98.5	97.7	98.2	98.4	98.3	96.6
EduBench	Content Relevance & Scope Control	98.8	97.8	99.1	-	99.3	98.7	98.6	99.2	99.2	95.9
EduBench	Basic Factual Accuracy	97.6	96.1	98	-	98.7	97.4	97.4	98.9	98.8	96.3
EduBench	Domain Knowledge Accuracy	94.1	94.5	98.7	-	97.1	94.8	97	96.3	97.3	95
EduBench	Clarity, Simplicity & Inspiration	88.7	88.7	96.4	-	91.9	89.2	93	90.9	92.9	90.9
EduBench	Higher-Order Thinking & Skill Development	78.8	76.1	91.2	-	81.6	78.2	86.1	76.9	83.6	80.4
EduBench	Scenario Element Integration	72.8	73.6	81.3	-	75.8	72.9	76	73.3	76.3	75.8
EduBench	Personalization, Adaptation & Learning Support	95.6	96.7	99.7	-	99.1	97.7	99.5	97.8	99.3	99.4
EduBench	Reasoning Process Rigor	95	95.8	98.6	-	97.5	94.4	96.4	96.9	98.3	92.8
Pedagogy Benchmark Multilingual	Average	87.72	88.59	97.28	88.48	87.17	86.52	82.5	92	92	89
Pedagogy Benchmark Multilingual	评估	90.13	91.48	97.76	91.03	89.24	87.89	85.65	92	92	89
Pedagogy Benchmark Multilingual	课堂管理	83.33	83.33	100	83.33	100	100	83.33	92	92	89
Pedagogy Benchmark Multilingual	教育理论	96.67	93.33	100	96.67	96.67	93.33	86.67	92	92	89
Pedagogy Benchmark Multilingual	学生理解	84.14	83.7	96.04	84.58	85.46	83.7	78.85	92	92	89
Pedagogy Benchmark Multilingual	教学策略	87.94	89.56	97.45	88.86	86.31	86.77	82.6	92	92	89

Usage

Transformers

python

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

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
Xuanhao Xie (谢轩豪)	RL training environment infrastructure; synthesis and training of general and education SFT data	Homepage
Shengyao Wang (王圣尧)	RL data processing and training for education-agent scenarios	GitHub
Yi Qian (钱毅)	General and education benchmark evaluation	GitHub
Jiahao Liu (刘家豪)	General and education benchmark evaluation	GitHub
Wenbo Wu (吴文博)	General-agent evaluation pipeline construction	Homepage

Citation

bibtex

@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.