SenseNova-SI-2B

The language backbone of SenseTime’s SenseNova-SI-1.1-InternVL3-2B spatial-intelligence model, packaged as a text-only GGUF for fast local inference with Ollama and llama.cpp.

🚀 Overview

SenseNova-SI is SenseTime’s open-source spatial-intelligence series, built on established multimodal foundations. The 2B member of the family, SenseNova-SI-1.1-InternVL3-2B, is a vision-language model fine-tuned from OpenGVLab/InternVL3-2B, which follows the ViT-MLP-LLM design and pairs an InternViT vision encoder with a Qwen2.5-1.5B language model.

This Ollama distribution is a text-only GGUF conversion of that model: it ships the Qwen2.5 language backbone extracted from the multimodal checkpoint. That is why the GGUF architecture metadata reports qwen2 and uses the ChatML prompt format, it is the LLM core of sensenova/SenseNova-SI-1.1-InternVL3-2B, without the vision encoder. If you need image understanding, use the full multimodal model upstream on Hugging Face; this build is for lightweight text-only chat and reasoning.

Two quantizations are published, both with a 32K context window, sized to run comfortably on laptops and edge hardware.

🎯 Key Features

• Tiny footprint, ~2B-parameter (Qwen2.5-1.5B backbone) GGUF; quants from ~1.0 GB run on modest CPUs and consumer GPUs.
• Long context, 32,768-token context window for multi-turn chat and longer documents.
• ChatML format, standard ChatML prompting, compatible with Ollama’s chat and OpenAI-compatible APIs.
• Spatial-intelligence lineage, derived from SenseTime’s SenseNova-SI series, which targets state-of-the-art spatial reasoning among open models of its size.
• Open license, Apache 2.0; free for research and commercial use.

🏷️ Available Versions

Two GGUF tags are published. Sizes are taken from the live Ollama page.

💻 Quick Start

Pull and run the recommended build:

ollama run richardyoung/sensenova-si-2b:Q4_K_M

Run the smaller quantization:

ollama run richardyoung/sensenova-si-2b:iq4_xs

OpenAI-compatible API:

curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "richardyoung/sensenova-si-2b:Q4_K_M",
    "messages": [{"role": "user", "content": "Describe the spatial relationship between a cup on a table and the floor."}]
  }'

🛠️ Use Cases

• Lightweight chat, fast conversational assistant on constrained hardware.
• Spatial-reasoning text tasks, describe and reason about spatial relationships in natural language.
• Edge & offline inference, runs locally without a GPU on the smaller quant.
• Prototyping, a compact backbone for experimenting with the SenseNova-SI line.
• Embedded assistants, small enough to bundle into resource-limited applications.

📋 System Requirements

A modern CPU runs both quants without a GPU; a GPU improves throughput. This is a text-only build, so no vision hardware or image preprocessing is required.

🔧 Technical Details

⚠️ Disclaimer

This is a text-only GGUF conversion of the language backbone of a vision-language model; it does not process images, despite the multimodal origin of its parent. It is the original, unmodified Qwen2.5 backbone redistributed for convenience and has not been abliterated or otherwise altered, retaining the alignment of the upstream release. The model is provided “as is” without warranty of any kind. Outputs may be inaccurate, biased, or otherwise unsuitable, verify important information independently and do not rely on the model for professional, legal, medical, or financial advice. You are responsible for ensuring your use complies with the Apache 2.0 license and all applicable laws.

🙏 Acknowledgments

• SenseTime / OpenSenseNova, for creating and open-sourcing the SenseNova-SI spatial-intelligence series.
• OpenGVLab, for the InternVL3 base model.
• Alibaba Qwen team, for the Qwen2.5 language backbone.
• llama.cpp / GGUF community, for the quantization and conversion tooling.
• Ollama, for the model runtime and distribution platform.

Built & maintained by Richard Young · DeepNeuro