Language: English | 中文

InnoSpark3.0-9B-260630

InnoSpark3.0-9B-260630 is an education-enhanced 9B model developed from Qwen/Qwen3.5-9B. The model is optimized for educational scenarios while preserving strong general capabilities in knowledge, reasoning, instruction following, and agent-style tasks.

InnoSpark3.0-9B-260630 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. Compared with the base model, we introduce 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 enhance visual capabilities; however, in our evaluation, compared with Qwen3.5-9B, the model shows only a slight visual-performance decrease of around 1%-2%.

Model Details

Item	Description
Model name	InnoSpark3.0-9B-260630
Base model	Qwen/Qwen3.5-9B
Parameters	9B
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:

1. 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.
2. 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 higher score in each row. Delta is calculated as InnoSpark3.0-9B - Qwen3.5-9B.

General Benchmarks

Type	Capability	Benchmark	Qwen3.5-9B	InnoSpark3.0-9B	Delta
Language	Knowledge	MMLU-Pro	81.71	83.2	1.49
Language	Knowledge	C-Eval	86.92	89.75	2.83
Language	Knowledge	SimpleQA-Verified	9.5	38.2	28.7
Language	Instruction Following	IF-Eval	77.26	86.88	9.62
Language	Instruction Following	IF-bench	48.81	77	28.19
Language	STEM & Reasoning	GPQA Diamond	82.32	81.31	-1.01
Language	STEM & Reasoning	LiveCodeBench v6	64.27	70.52	6.25
Language	STEM & Reasoning	AIME25	50	93.33	43.33
Language	STEM & Reasoning	AIME26	50	93.33	43.33
Language	Coding Agent	SWE-bench Verified-Agentic	41.4	41.4	0
Language	Coding Agent	Terminal-Bench 2.1	20.22	24.56	4.34
Language	General Agent	BFCL_v4	61.86	74.54	12.68
Language	General Agent	TAU3-bench	61.86	64.8	2.94
Vision-Language	STEM & Puzzle	MMMU-Pro	76.71	75.84	-0.87
Vision-Language	General VQA	MMBenchEN-DEV-v1.1	91.23	90.45	-0.78
Vision-Language	Document Understanding	OCRBench	87.6	85.1	-2.5

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	Qwen3.5-9B	InnoSpark3.0-9B	Delta
EduBench	Total	84.8	87.4	2.6
EduBench	Instruction Following & Task Completion	94.8	98.6	3.8
EduBench	Role & Tone Consistency	87.7	87.8	0.1
EduBench	Content Relevance & Scope Control	95.8	97.9	2.1
EduBench	Basic Factual Accuracy	91.1	94.6	3.5
EduBench	Domain Knowledge Accuracy	84.8	90.1	5.3
EduBench	Clarity, Simplicity & Inspiration	83	84	1
EduBench	Higher-Order Thinking & Skill Development	74.2	69.4	-4.8
EduBench	Scenario Element Integration	43	56.8	13.8
EduBench	Personalization, Adaptation & Learning Support	89.4	94.8	5.4
EduBench	Reasoning Process Rigor	92.3	92.9	0.6
Pedagogy Benchmark Multilingual	Assessment	73.54	85.65	12.11
Pedagogy Benchmark Multilingual	Classroom Management	83.33	100	16.67
Pedagogy Benchmark Multilingual	Educational Theory	66.67	83.33	16.66
Pedagogy Benchmark Multilingual	Student Understanding	60.35	81.5	21.15
Pedagogy Benchmark Multilingual	Teaching Strategy	70.53	86.08	15.55
Pedagogy Benchmark Multilingual	Total	68.7	84.78	16.08

Usage

Transformers

import torch
from transformers import AutoProcessor, AutoModelForMultimodalLM

model_id = "sii-research/InnoSpark3.0-9B-260630"

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/InnoSpark3.0-9B-260630 --trust-remote-code

Intended Use

InnoSpark3.0-9B-260630 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, InnoSpark3.0-9B-260630 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

@misc{innospark3_9b_260630,
  title = {InnoSpark3.0-9B-260630},
  author = {SII Research},
  year = {2026},
  howpublished = {\url{https://huggingface.co/sii-research/InnoSpark3.0-9B-260630}}
}

Please also follow the citation and license requirements of the base model Qwen/Qwen3.5-9B.