GitHub radar

Run an open model on your PC that thinks deeper than its size

SenseNova-Loop-9B is a free model for reasoning, coding, tool use and image tasks. You can run it on your own computer if your graphics card has enough memory.

01sensenova/SenseNova-Loop-9B-GGUF 17181k downloads/mo9B paramsimage-text-to-text

It is an open model with 9 billion parameters. It runs its middle layers twice, so it computes more deeply than a model of this size usually does. It reads text and images and replies in text.

When it helps

It is useful if you have a strong graphics card and want to test a local model on reasoning, code or images. For serious work that has to be reliable, choose a proven tool.

Pros

  • Apache 2.0 license: free to use
  • Understands both text and images
  • The 4-bit version takes about 4.9 GB
  • Reads up to 512,000 tokens (word pieces) at once

Cons

  • Needs a graphics card with about 10 GB of memory, no Mac figures
  • Needs the authors' own build of llama.cpp
  • All benchmark scores come from the authors and are not independently checked

How to set it up — step by step

  1. 1Check how much memory your graphics card has. The 4-bit version needs about 10 GB.
  2. 2Send your agent the text from the block below. It will explain what makes the model unusual and check your memory.
  3. 3Ask your agent to build llama.cpp (a program that runs models) from the OpenSenseNova/llama.cpp repository on GitHub.
  4. 4Start the model with llama serve -hf sensenova/SenseNova-Loop-9B-GGUF:Q4_K_M -ngl 99 -c 4096 --port 8000.
  5. 5Give the model a simple task and check the answer. The authors warn that its coding and tool use can be uneven.

Text for your agent

Copy this and send it to your agent — Claude Code, Codex, any of them:

Open https://huggingface.co/sensenova/SenseNova-Loop-9B-GGUF, explain in plain words how this model differs from ordinary 9-billion-parameter models, and tell me whether my computer has enough memory to run the 4-bit version through their llama.cpp build.

A regular laptop is not guaranteed. The authors estimate about 10 GB of graphics card memory for text up to 64 thousand tokens and about 15 GB for 256 thousand, with 4-bit weights. The 4-bit weights themselves take about 4.9 GB. The card gives no figures for a Mac.

Open on Hugging Face