What are Meta's Muse models, and are they any good for coding?

Meta started the Muse family in April 2026 with Muse Spark, and has shipped quickly since. By late October there are four names a programmer should know, and they come in two kinds: hosted models, which run on Meta's servers and which you pay for per use, and open-weight models, whose files you can download and run on hardware you control.

For coding, Meta offers its own agent, Muse Code, a command-line tool for macOS and Linux that reads your project, edits files and runs commands, much like Claude Code or Codex.

The line-up as of October 2026

Prices come from Meta's API pricing page; a token is a short piece of text, about three-quarters of a word.

What Meta says about coding

Meta presents Muse Spark 1.3 as a step up in agentic and coding work. It says the model was trained on more long coding tasks, asks clarifying questions when a request is vague, asks for help when it is stuck, and confirms before doing anything with lasting effects. In Meta's internal comparisons it used about 20% fewer tool calls and about 25% fewer tokens than Spark 1.2, and wrote less padding. Those are useful traits in an agent, because every extra step costs time and money.

Muse Wonder is Meta's most capable model, built for personal and professional agent tasks. It costs the same per token as Claude Opus 5.5.

What outside tests show

Independent results are thinner than for Claude or GPT, and so far less flattering. Google's Android Bench, a public test of long coding tasks, does not include Muse models yet. Vals AI, an independent testing firm, ran Muse Spark Open on Terminal-Bench 4.0, a test of multi-step command-line work, and it solved 6.6% of tasks, far behind the hosted leaders in the 55% to 65% range on that firm's runs. That figure belongs to the open model, not to the hosted Spark 1.3 or Wonder, but it is a reminder that a launch chart is not the whole picture.

Muse Glimmer has strong numbers for its size on Meta's own card: 51.2% on SWE-bench Pro, a test of real software fixes, and 76.0% on the older SWE-bench Verified. Those are vendor figures, and both tests differ from the ones the big labs now quote.

A fair way to judge any Muse model is to give it a task you have already solved:

This repository has a failing test in tests/test_invoice.py. Find the cause, fix it in the smallest change you can, run the full test suite, and explain the fix in three sentences. Ask me before adding any dependency.

Which Muse model can you run yourself?

For most people, the answer is Muse Glimmer-30B. It has about 29.6 billion parameters, the numbers a model learns during training, and Meta built it for agent work on consumer hardware. Meta's 4-bit build, which stores each parameter in roughly half a byte, is a 16.8 GB file meant for a graphics card with 24 GB of memory; a slightly better 19.7 GB build targets 32 GB cards. It reads up to 131,072 tokens of context, accepts text and images, and runs in llama.cpp (build b10353 or newer). Everything Meta released for it is under Apache 2.0, so you may use it commercially. The step-by-step setup is covered in the guide to running Muse Glimmer locally.

Muse Spark Open is a different animal. Its name, 816B-A42B, follows the usual convention for a mixture-of-experts model: 816 billion parameters in total, with about 42 billion active for each token. Even at half a byte per parameter, the weights alone would take roughly 400 GB, which means data-centre hardware, not a laptop. Meta's model page points to server tools such as vLLM and SGLang. The licence also deserves a careful read: it allows use, fine-tuning and redistribution, but any product built on it must display "Built with Muse Spark Open", and companies that sell model access with more than $20 million in yearly revenue must sign a separate agreement with Meta.

Muse Spark 1.3 and Muse Wonder have no downloadable weights, so they run only through Meta's services. If keeping your code on your own machine is the goal, Glimmer is the Muse model to start with.