Meta Open-Sources Its Biggest Model Yet

On Thursday, Meta did something few people expected it to do this year: it released the weights for Muse Spark Open, a model it describes as the most capable open-weight AI model built in the United States. The model has 816 billion total parameters, with 42 billion active on any given token. It is natively multimodal, supports a one-million-token context window, and was built for coding and long-horizon agentic work. You can download it from Hugging Face today, and hosted APIs are already live on Groq, Fireworks, Together, and OpenRouter.

This is a reversal. When Meta launched Muse Spark in the spring, it kept the weights closed. Now it has gone the other way, and the timing suggests Meta has decided the open strategy is worth more than the control.

The biggest surprise is not the release. It is who helped build it.

What Meta actually released

816 billion parameters sounds enormous, and it is, but the number that matters more is 42 billion. That is the count of parameters active for each token, which means the model works like a team of specialists where only the relevant experts wake up for each word. That is the standard trick that lets giant models stay affordable to run, and it is why Meta can talk about frontier-class performance without pretending you can run the full model on a laptop.

One million tokens of context is roughly 750,000 words. That is a small library of text in a single prompt. For coding and agentic work, where the model has to hold an entire codebase or a long chain of tool calls in its head, context like that is a genuine advantage.

Early independent benchmarks cited in coverage put Muse Spark Open in an interesting spot: roughly even with Mistral Large 4 on general tasks, ahead of NVIDIA’s Nemotron 3 Ultra on agent work, but behind the top Chinese open models on coding. The exact rankings will shift as more people test the model, but the shape of the field is clear. The open-model race now has a serious American contender again, and it is fighting hardest on agentic tasks, which is where the industry’s money is moving.

Why Meta reversed course

Six months ago, Muse Spark launched without open weights. Meta’s open strategy, the thing that made Llama famous and turned “open” into Meta’s brand advantage, appeared to be on hold. The closed launch suggested Meta had decided the future belonged to the big proprietary models and their APIs.

Muse Spark Open is an admission that the math changed. Open weights buy Meta something money cannot: a developer ecosystem that builds on its models for free. Every startup that fine-tunes Muse Spark Open, every research lab that probes it, every hobbyist who runs it locally is working inside Meta’s orbit. The Llama era proved this playbook works. Meta seems to have decided that whatever it gains from keeping weights closed is smaller than what it loses.

There is also a competitive reason. Chinese labs have been releasing strong open models at a steady pace, and several of them lead on coding benchmarks. Releasing an 816-billion-parameter open model is Meta planting its flag and saying the United States can play this game too.

The Apple alliance is the real surprise

Buried in the release is the detail that will matter most to practitioners: Meta worked with Apple for months to optimize the model for MLX, Apple’s machine-learning framework. A quantized version can run locally across two to four M5 Ultra Mac Studios pooling roughly a terabyte of unified memory.

Read that twice. A frontier-class model from Meta, running on Apple hardware, in your own office. Five years ago, the idea of Meta and Apple cooperating on AI infrastructure would have sounded like a joke. The old rivalries are being rewritten by the economics of this race. Apple has the silicon and the memory bandwidth; Meta has the model weights. Neither could offer this setup alone.

The setup is not cheap. Two to four top-end Mac Studios is a serious hardware purchase. This is self-hosting for well-funded teams, not weekend projects. But the fact that the option exists at all changes the conversation about who controls frontier AI. You no longer have to choose between a frontier model and keeping your data in your own building.

The fine print: license and hardware

Two caveats will decide how far this release goes.

The first is the license. The weights come under the Muse Model License, which has commercial strings attached: attribution is required, and commercial terms kick in above $20 million in revenue. That is not a fully open license. Big companies will need to read the fine print and possibly negotiate. For startups and researchers, the terms are workable. For a company approaching scale, the license is a decision point, not a formality.

The second is the hardware bar. Even quantized, this is a multi-machine model. Running it locally means committing real capital to Apple hardware. Most users will hit Muse Spark Open through the hosted APIs on Groq, Fireworks, Together, and OpenRouter, which is exactly how the ecosystem has absorbed every other big open release. Self-hosting is the headline; the APIs are the distribution.

Where it sits in the open-model race

Zoom out, and the pattern is hard to miss. The open-weight frontier is splitting into two camps: American labs releasing big models with commercial license caveats, and Chinese labs releasing strong models that lead on benchmarks like coding. Meta’s release does not end that contest, but it keeps the American side in the fight. For the first time in a while, the strongest open-model story of the year comes from the United States.

There is also a feedback loop to watch. Open releases get probed, red-teamed, and fine-tuned by thousands of people within weeks. That scrutiny finds weaknesses and strengths faster than any internal testing program. Meta is buying that scrutiny for free, and it makes the next model better.

What it means for builders

If you build with AI, three things change this week.

First, you have a new self-host option for frontier-class agentic work. If your customers care about data staying inside their walls, or your margins are being eaten by API bills, Muse Spark Open is worth evaluating. The license terms above $20 million in revenue are the question to answer first.

Second, the API market gets more competitive. Groq, Fireworks, Together, and OpenRouter all hosting the model on day one means inference pricing pressure on everyone else. Cheaper inference is good for builders and bad for anyone whose moat is “we host a model.”

Third, the Apple angle deserves a second look. If serious AI workloads run well on Mac Studios, the hardware story of AI gets more interesting. Not everyone who needs local inference wants a rack of GPUs. A desk-side cluster of Macs running a frontier model is a new kind of computing setup, and Apple clearly intends to be in this market.

What to watch next

Three questions will decide whether Muse Spark Open is a milestone or a moment.

One: do the benchmarks hold up? Early independent results favor it on agent tasks and trail the Chinese models on coding. The next few weeks of public testing will settle the real ranking.

Two: who adopts it? The license’s $20 million revenue line is a cliff that will shape adoption. Watch for startups and research labs first, enterprises after legal reviews.

Three: how do the Chinese labs respond? The open-model race has been theirs to lead for months. Meta just raised the stakes. The next release from that side will tell us whether this is an arms race or a one-sided surge.

The open era of AI had a quiet year. It just got loud again.

Sources

• MacRumors, October 15, 2026
• Reuters, October 15, 2026
• Flavio Copes, October 15, 2026

Related reading

• Meta’s Muse Wonder Is Its First Real Shot at the AI Frontier
• Anthropic Just Signed a Safety Pact With Singapore. Governments Are Becoming AI Infrastructure