Table with columns: Item, Description| Item | Description |
|---|
| Model name | InnoSpark3.2-27B-1102 |
| Base model | Qwen3.8-27B |
| Parameters | 27B |
| Post-training method | Reinforcement learning (RL) |
| Post-training focus | Educational QA, tutoring, educational assistance, and multi-turn InnoAgent training |
| Primary use cases | Educational QA, tutoring, educational assistance |
| Evaluation benchmarks | EduBench, LongTutor |
| Agent evaluation | InnoBench |
RL Post-Training
InnoSpark3.2-27B-1102 is based on Qwen3.8-27B and has undergone RL post-training for educational question answering, tutoring, and educational assistance.
Compared with the previous release, InnoSpark3.1-27B-1008, the main update in InnoSpark3.2-27B-1102 is stronger multi-turn RL training for InnoAgent, a personalized teaching agent. Training focuses on knowledge explanation, learning diagnosis, exercise generation and reuse, and learning planning, aiming to improve response quality, tool use, and learner profile maintenance in multi-turn interactions.
Evaluation Results
EduBench: Scores by Dimension
Table with columns: Dimension, Qwen3.8-27B, InnoSpark3.2-27B-1102, Change| Dimension | Qwen3.8-27B | InnoSpark3.2-27B-1102 | Change |
|---|
| Instruction Following & Task Completion | 9.608 | 9.847 | +0.239 |
| Content Relevance & Scope Control | 9.715 | 9.905 | +0.190 |
| Scenario Element Integration | 8.434 | 9.241 | +0.807 |
| Basic Factual Accuracy | 9.522 | 9.765 | +0.243 |
| Domain Knowledge Accuracy | 9.049 | 9.610 | +0.561 |
| Reasoning Process Rigor | 9.434 | 9.828 | +0.394 |
| Clarity, Simplicity & Inspiration | 8.788 | 9.044 | +0.256 |
| Higher-Order Thinking & Skill Development | 8.332 | 9.451 | +1.119 |
| Personalization & Learning Support | 9.375 | 9.973 | +0.598 |
| Role & Tone Consistency | 9.008 | 9.936 | +0.927 |
* EduBench evaluates three task types, IP, PLS, and TMG, using 3,025 Chinese samples. Scores are out of 10, and changes are calculated from unrounded values.
LongTutor
Table with columns: Metric, Qwen3.8-27B, InnoSpark3.2-27B-1102, Change| Metric | Qwen3.8-27B | InnoSpark3.2-27B-1102 | Change |
|---|
| Overall memory accuracy | 87.20% | 88.90% | +1.70 pp |
| Information extraction accuracy | 97.80% | 98.30% | +0.50 pp |
| Multi-session reasoning accuracy | 72.90% | 74.40% | +1.50 pp |
| Hallucination check accuracy | 90.90% | 94.00% | +3.10 pp |
| Diagnosis accuracy | 42.20% | 43.80% | +1.60 pp |
| Diagnosis Macro-F1 | 0.3738 | 0.4084 | +0.0346 |
| Use of learning history | 3.828 | 4.366 | +0.538 |
| Strategy consistency | 4.352 | 4.845 | +0.493 |
| Coherence | 4.311 | 4.675 | +0.364 |
| Ease of understanding | 4.058 | 4.307 | +0.249 |
| Overall teaching quality | 4.038 | 4.519 | +0.481 |
* LongTutor is evaluated on 1,000 samples from the human-annotated XES3G5M Gold set. Teaching scores are out of 5, and changes are calculated from unrounded values.
InnoBench: Native Agent Environment Evaluation
InnoBench is a benchmark for evaluating task capabilities within InnoAgent, a personalized teaching agent. It covers knowledge explanation, learning diagnosis, exercise generation and reuse, learning planning, and learner profile maintenance. Evaluation runs in the InnoAgent environment and examines interactions, tool execution results, and state changes to assess the model's ability to understand teaching needs, use tools, and complete tasks.
Table with columns: Selected task (strict success rate), Qwen3.8-27B, InnoSpark3.2-27B-1102| Selected task (strict success rate) | Qwen3.8-27B | InnoSpark3.2-27B-1102 |
|---|
| Knowledge explanation | 33.33% | 100.00% |
| Diagnostic feedback | 33.33% | 43.33% |
| Clarification and planning | 0.00% | 66.67% |
| Exercise reuse | 20.00% | 100.00% |
| Knowledge retrieval | 93.33% | 100.00% |
| Knowledge archiving | 44.44% | 100.00% |
Usage
import torchfrom transformers import AutoProcessor, Qwen3_5ForConditionalGeneration model_id = "sii-research/InnoSpark3.2-27B-1102" processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)model = Qwen3_5ForConditionalGeneration.from_pretrained( model_id, torch_dtype="auto", device_map="auto", trust_remote_code=True,) messages = [ { "role": "user", "content": [ { "type": "text", "text": "Explain the quadratic formula in a way that is suitable for middle school students, and give an example problem.", } ], }] 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/InnoSpark3.2-27B-1102 --trust-remote-code
Intended Use
- Educational QA: explain concepts, answer subject-specific questions, and discuss approaches to solving problems.
- Tutoring: guide problem solving through questions, provide support based on learning history, and adapt explanations.
- Educational assistance: support lesson planning, teaching material generation, and instructional strategy recommendations.
- InnoAgent integration: serve as the model backend for InnoAgent, combining tool use and learner profile maintenance to support multi-turn teaching tasks such as knowledge explanation, learning diagnosis, exercise generation and reuse, and learning planning.
Limitations
Like other large language models, InnoSpark3.2-27B-1102 may generate inaccurate, incomplete, or biased content. In high-stakes learning, assessment, or student guidance settings, model outputs should be reviewed by qualified educators. The model should not be the sole basis for fact-checking, grading decisions, psychological counseling, medical advice, legal advice, or other safety-critical decisions.
Evaluation results may be affected by prompt format, decoding parameters, evaluation implementation, and data version. Users should conduct additional evaluations for their specific use cases before deployment in production or real classrooms.
Contributors
License
See LICENSE for the Apache-2.0 license included with this checkpoint.
Citation
@misc{innospark3_2_27b_1102, title = {InnoSpark3.2-27B-1102}, author = {Siyu Song and Shengyao Wang and Siyuan Su and Ye Lu}, year = {2026}}