wasl.tech

Meta Muse Spark Open: 816B Open-Weight AI Model for Agentic Coding

Picture a developer at 2 a.m., staring at a bug she cannot reproduce. She has a screenshot of the error, a ten-second voice note from a teammate describing what happened, and a screen recording of the crash.

Until now, she would have fed those into three different tools. Meta’s answer, released on November 17, is called Muse Spark Open — one model that takes all three at once, and Meta has put its weights on Hugging Face for anyone to download.

The release is the biggest open-weights move Meta has made since it rebuilt its AI division from scratch. Muse Spark Open is an 816-billion-parameter model designed for agentic coding, and Meta describes it as the most capable United States open-weights model released so far.

What Muse Spark Open Actually Is

Muse Spark Open is the full-size model in Meta’s new Muse series, the family built by Meta Superintelligence Labs. The division was set up nine months ago to rebuild Meta’s AI stack from the ground up, under former Scale AI chief Alexandr Wang.

The first Muse model, the proprietary Muse Spark, now powers the Meta AI assistant at meta.ai and in the Meta AI app. A smaller dense model, Muse Glimmer, followed in August for running locally.

Muse Spark Open is the open-weights member of the family. It is built for one job above all: agentic coding. It reads a codebase, writes and debugs code, calls tools, and works through long, multi-file jobs on its own.

Meta says it trained the model on real developer workflows, including multi-file refactors and long debugging sessions. It plans, acts, observes what happened, and then corrects itself — the loop that separates an agent from a chatbot that just answers questions.

816 Billion Parameters, Only 42 Billion at Work

The headline number is 816 billion parameters. The number that matters more is 42 billion, because that is roughly how much of the model actually computes each token.

Muse Spark Open is a mixture-of-experts model. Each of its 62 layers contains 256 expert networks, and a router picks just eight of them for every token. The experts hold about 97 percent of the model’s parameters.

The result is a model that runs at the speed of a 42-billion-parameter system while keeping the knowledge of an 816-billion-parameter one. That is the whole point of the design: frontier-scale capability without frontier-scale compute on every single token.

There is a catch, and Meta does not hide it. All 256 experts stay in memory even though only eight run at a time. Deployment needs the memory of an 816-billion-parameter model, and the full-precision checkpoint is about 1.5 terabytes.

Muse Spark Open Sees, Hears and Watches While It Codes

Most coding models read text and nothing else. Muse Spark Open also takes images, audio and video as input, which is what makes the 2 a.m. debugging scene possible.

A perception encoder with 1.8 billion parameters converts images and video frames into tokens the language model can reason over. The same model can read a screenshot of a broken layout, watch a screen recording of a crash, or listen to a voice note describing the problem.

Audio goes through its own encoder, producing one token every 80 milliseconds. Video is sampled at two frames per second by default, with timestamps woven into the text so the model can connect what it sees to when it happened.

For a developer, that means one model can look at the error, listen to the report and read the code, instead of three tools passing notes to each other.

A Million Tokens of Working Memory

Muse Spark Open carries a context window of one million tokens. That is enough room to hold an entire project — complex dependencies, thousands of files — inside a single conversation.

Long context matters most for agentic coding, where the model needs to keep the whole codebase in mind while it plans a refactor across dozens of files. A model that forgets the edges of a project cannot safely change its middle.

The model also exposes a reasoning effort control, with levels from low to xhigh. Reasoning cannot be switched off entirely, so users who want fewer tokens on an easy task keep it at low.

Running Muse Spark Open: From a Mac to a GPU Cluster

Meta ships the model with deployment recipes for llama.cpp, vLLM, SGLang and MLX, plus guidance for running and fine-tuning it. The checkpoints live on Hugging Face, including a 4-bit NVFP4 version built for NVIDIA Blackwell GPUs.

The full-precision checkpoint runs across four nodes of eight H200 GPUs. The NVFP4 version fits on a single B200 node, and SGLang’s reference setup runs it across four GB300 GPUs with an OpenAI-compatible API.

Quantized versions bring it within reach of very large local machines. Meta tested a 2-bit llama.cpp build on a Mac M3 Ultra with 512 gigabytes of memory, where it ran at 18 tokens per second while using 304 gigabytes. An MLX build on the same machine managed about 21 tokens per second.

Nobody will call that fast, but it means a frontier-class coding model can now sit on a desk instead of in a data centre.

NVIDIA is leaning in too. The model was trained on GB300 NVL72 Blackwell Ultra systems, and NVIDIA offers a one-command NIM container for production deployment plus its NeMo AutoModel library for full fine-tuning and LoRA adaptation without converting checkpoints.

Why Meta Is Opening Up Its Biggest Coding Model

Meta’s original Muse Spark is proprietary, and it only opened a private API preview to selected users. Muse Spark Open is the opposite bet: give the weights away and let developers build on top of them.

The move lands in the middle of an open-weights race. Meituan released its LongCat-2.5 model under the permissive MIT licence earlier this month, and Chinese labs have been shipping large open models all year.

Meta, which once defined open AI with Llama, has spent the past year watching rivals take that ground. Spark Open is its answer: a United States model, built for the most commercially valuable job in AI right now, available to download and fine-tune.

The release also follows the launch of the Muse Spark 1.1 API on November 18, Meta’s first hosted model API for outside developers. Meta is now selling access to the closed model and giving away the open one at the same time.

The Catch: 816 Billion Parameters Need a Home

For all the talk of openness, Muse Spark Open is not a model most people will run. The weights are free, but the hardware is not, and the 1.5-terabyte checkpoint is a reminder that sparse compute still means dense memory.

Meta’s own claims — higher first-attempt accuracy, reliable tool calling, deep understanding across repositories — are vendor figures with no independent verification yet. Benchmarks from outside labs will decide whether the model is genuinely frontier-class or merely frontier-sized.

There is also the question of what “open” means here. Meta has published weights, recipes and fine-tuning guidance, but the training data stays private, and the licence terms will decide how freely companies can build products on top of it.

What Happens Next for Muse Spark Open

The first test will be what developers actually build with it. Watch for fine-tuned versions aimed at specific stacks, and for coding agents that feed it screenshots and screen recordings instead of pasted error messages.

The second test is benchmarks. Independent evaluations of its coding accuracy, tool use and long-context reliability will show whether Meta’s claims hold up against rivals like Gemini 4 Carbon and Qwen4-Max, and against open-weights models such as LongCat-2.5.

Meta’s own roadmap points further up the ladder. The proprietary Muse Spark already runs a multi-agent “Contemplating” mode for extreme reasoning, and Meta has said it hopes to open-source future versions of the model.

If Spark Open is the first step, the next ones will be judged on a simple question: does an open model this large actually help developers ship better software, or does it just move the hardware bill from Meta’s data centres to theirs?

For now, the weights are out and the recipes are published. The 2 a.m. debugger has a new option: paste the screenshot, attach the voice note, drop in the screen recording — and let one model sort through all three.

According to a detailed technical breakdown of the release, that multimodal, million-token design is exactly what Meta built it for.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top