Meta's Muse Spark Open: A Giant Model Anyone Can Download

Meta has released Muse Spark Open, an open-weight version of its flagship AI model that anyone can download for free, and it is one of the biggest open model drops of the year. The company published the weights on October 15, turning what had been a closed, API-only system into something developers can run, fine-tune, and build on themselves.

The model is enormous. It is built as a mixture-of-experts system with more than eight hundred billion total parameters, of which around forty-two billion activate for each token it processes. It handles a million-token context window, meaning it can work with entire codebases or long documents in a single pass, and it accepts text, images, audio, and video as input while producing text. Meta built it for coding and agentic work: the kind of long, tool-using tasks where a model plans, calls software, and keeps going. Meta's chief of AI, Alexandr Wang, called it the most capable US open-weight model to date, according to the company's announcement, which was covered in detail by Particle News.

Why does this matter? For the past year the open-model leaderboard has been dominated by Chinese labs. Models from teams behind Qwen, Kimi, and GLM set the pace for downloadable weights, and American open releases lagged behind. Muse Spark Open is the West's first heavyweight answer at this scale, and it is natively multimodal where several of its rivals are text-only. The reaction from the open-model community was instant. "I waited so long for US open weight models to be relevant again," wrote Interconnects author Nathan Lambert on X, adding: "Still more work to do!"

One important distinction: open-weight is not the same as open-source. Meta published the weights but not the training data or training code, and the release sits under a custom Muse Model License rather than a permissive open-source license. As one developer breakdown on Dev.to put it, you can download it today, but you probably cannot run it.

What the numbers actually say

The benchmark story comes with a large caveat: every performance number published so far comes from Meta or its launch partners, and none had been independently confirmed at launch. That said, the company's own scorecard shows a mixed picture. On the Terminal-Bench coding test, the model scored seventy-six point four, trailing Kimi K3 at eighty-five, GLM 5.3 at eighty-three point nine, and Qwen3.8 at eighty-two. On the independent Artificial Analysis Intelligence Index it landed at thirty-nine against GLM 5.3's forty-five.

Where it leads is tool use and agentic tasks. On the MCP-Atlas test, which measures how well a model works with external tools, it scored eighty-nine point two against eighty-four point two for GLM 5.3, and it also came out ahead on JobBench and the hard-question HLE-Diamond set. The pattern is clear: it trails the best Chinese open models at raw coding, but wins at orchestrating tools and agents. If your workflow is a model driving software rather than just writing code snippets, that trade-off favors Meta.

The catch: hardware, license, and who can actually use it

Running the full model locally is a serious undertaking. Meta's model card asks for at least seven hundred gigabytes of graphics memory for the compressed checkpoint, or two terabytes for the full-precision version. On the consumer-hardware side, Meta worked with Apple to run it through the MLX framework across two linked Mac Studios fitted with the top-end M5 Ultra chip, reaching roughly twenty-five tokens per second. The Studio configuration needed to join that pair starts at a price above ten thousand dollars, so this is workstation territory, not laptop territory. Quantized versions shrink the footprint to around three hundred gigabytes, which is still far beyond what a normal machine holds.

The license adds a second gate. Products built with the model must display the words "Built with Muse Spark Open," and any company with more than twenty million dollars in annual revenue needs a separate agreement with Meta before it can sell access to the model as a hosted service. Military and critical-infrastructure uses are barred outright. Training data and training code stay private, so researchers can probe the model's behavior but not fully reproduce how it was made.

So who should actually care? If you build agentic coding workflows and want an open-weight model with a million-token window and multimodal input, this is worth your time. The cheapest way to try it is a hosted API endpoint through one of the day-one providers like Together AI, Fireworks, Groq, or OpenRouter, where Together lists pricing at a dollar and a quarter per million input tokens and four dollars and a quarter per million output tokens. Skip it if you wanted a model for your laptop, if you need a permissive commercial license without attribution strings, or if you already have a Qwen or Kimi-class model doing the job.

How we got here

The Spark family has moved fast. Meta introduced the first proprietary Muse Spark model in April, expanded it with version 1.1 in July, and iterated through versions 1.2 and 1.3 across August and September with better long-horizon autonomy and multi-agent orchestration. In mid-October the company also launched Muse Wonder, a closed frontier model. Spark Open, released October 15, is the open-weight sibling: related lineage, different access model. Until now, the only open-weight Muse model was a small thirty-billion-parameter release called Glimmer.

The big picture is a shift in the open-model race. The frontier went natively multimodal with this release, and the United States finally has a downloadable model at the same scale as the Chinese open weights it has been chasing. The open question, literally, is verification: until independent labs publish their own numbers, every claim about this model is Meta's. The weights are out there now, so those independent tests are coming. For a generation of builders deciding what to build on, the next few weeks of third-party results will matter more than the launch-day scorecard. That is also the pattern across AI right now: partnerships and splits are reshaping who controls the models, while agentic AI is moving into real jobs like PR and AI glasses are changing how people experience the physical world.