---

license: apache-2.0

base_model:

• AI4SGI/ExoMind-9B

library_name: transformers

tags:

• text-generation-inference
• llama-cpp
• qwen3_5
• exomind
• scientific-reasoning
• scientific-research
• agentic
• tool-use
• multimodal
• vision-language
• safetensors

language:

• en
• zh

pipeline_tag: text-generation

---

AI4SGI-ExoMind-9B-GGUF

> ExoMind-9B is the compact checkpoint in Shanghai AI Laboratory's ExoMind family, fine-tuned from Qwen3.5-9B for lower-resource experimentation in scientific reasoning and agentic research under the same "extended-mind-inspired" paradigm as the larger 35B-A3B flagship — organizing the model, specialized interaction objects, and autonomous interaction processes into one unified system. It's trained via progressive Chain-of-Interaction (CoI) training on selected pure-reasoning and interaction trajectories, enabling workflows around source discovery, evidence grounding, executable verification, and observation integration for scientific inquiry, while retaining the native image-text multimodal capabilities of its Qwen3.5 base and supporting a 262,144-token context window. Formal benchmark scores are reported only for the main ExoMind-35B-A3B system — which posts leading results among frontier models on tasks like FrontierScience-Research (70.0), CMT-Benchmark (84.0), AMO-Bench (78.0), and an eight-benchmark scientific reasoning average of 68.3, outperforming Claude-Opus-4.8 Thinking, GPT-5.5, and Gemini-3.1-Pro Preview — while ExoMind-9B itself has not been separately scored and is instead positioned as a resource-conscious checkpoint for scientific question answering, mathematical/computational reasoning, tool-use experiments, and agentic prototyping. It's servable via vLLM or SGLang with Qwen3-style reasoning and tool-call parsers, available in official and community GGUF quantizations, and released under the Apache License 2.0 (with the accompanying preprint, figures, and ExoMind branding separately governed by dedicated content and brand terms).

Model Files

File Name | Quant Type | File Size | File Link |

|-----------|------------|-----------|-----------|

| ExoMind-9B.BF16.gguf | BF16 | 17.9 GB | Download |

| ExoMind-9B.Q3_K_L.gguf | Q3_K_L | 4.93 GB | Download |

| ExoMind-9B.Q3_K_M.gguf | Q3_K_M | 4.62 GB | Download |

| ExoMind-9B.Q3_K_S.gguf | Q3_K_S | 4.26 GB | Download |

| ExoMind-9B.Q4_0.gguf | Q4_0 | 5.31 GB | Download |

| ExoMind-9B.Q4_K_M.gguf | Q4_K_M | 5.63 GB | Download |

| ExoMind-9B.Q4_K_S.gguf | Q4_K_S | 5.35 GB | Download |

| ExoMind-9B.Q5_0.gguf | Q5_0 | 6.31 GB | Download |

| ExoMind-9B.Q5_K_M.gguf | Q5_K_M | 6.47 GB | Download |

| ExoMind-9B.Q5_K_S.gguf | Q5_K_S | 6.31 GB | Download |

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp