GitHub radar

Muse Spark Open: Meta's open source model for code and agents

Meta released Muse Spark Open, an open source model built for code and agentic tasks. You can try it free in your browser, no hardware needed.

01meta-models/Muse-Spark-Open-816B-A42B 6675.2k downloads/mo816B paramsimage-text-to-text

A large language model from Meta with open weights: 816 billion parameters, but only 42 billion are used per token (a word fragment) — that's the efficiency trick. Built for code and agentic tasks, it also understands text, images, video, and audio, but replies only in text.

When it helps

Good if you want to test an agent for terminal or coding tasks without buying hardware. Not useful if you want to self-host — it needs data-center-grade hardware.

Pros

  • Free to try in your browser via HuggingChat
  • Built specifically for code and agentic tasks
  • Understands text, images, video, and audio
  • Free commercial use until $20M in yearly revenue

Cons

  • Self-hosting is nearly impossible: needs a 700GB-VRAM GPU
  • Terminal-Bench 2.1 score is self-reported, not independently verified

How to set it up — step by step

  1. 1Open HuggingChat at https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B — no hardware required.
  2. 2Send the agent the text from the block below.
  3. 3Give the model an agentic task in a terminal: ask it to write and run code for a specific task.
  4. 4Check how the model reasons and what code it produces before trusting it with a real task.
  5. 5If you want to self-host, note it needs a 700GB-VRAM GPU or two Mac Studios with M5 Ultra chips.

Text for your agent

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

Open HuggingChat and try Meta's Muse Spark Open model: https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B — give it an agentic terminal task and show me how it reasons and what code it writes.

Server-grade hardware - you cannot run this at home: it needs a GPU with 700 GB of VRAM in the compressed NVFP4 version, or two Mac Studios with M5 Ultra chips through MLX. You can try it for free in HuggingChat, without any hardware of your own.

Open on Hugging Face