Meta Puts America Back in the Open Weights Race With Muse Spark Open

With the launch of Muse Spark Open, an 816 billion parameter model, Meta has put forward what it terms the most able US open weight model so far. To show off its mettle, the company had the model running locally on a pair of Apple Mac Studios. It is a change of pace Silicon Valley has been looking for after a year where the open model charts were the province of Chinese labs. Now the matter is whether the industry at large can make use of it given the license and hardware costs as well as the gaps in the benchmarks.

The Biggest US Open Model Yet, With No Data Center Required

Meta Superintelligence Labs put out Muse Spark Open on October 15, accompanied by a launch post and a model card on Hugging Face. A mixture of experts in design, the model has 816 billion parameters in all, with 42 billion active per token. It will take in text, video, images and audio but is limited to generating text (VentureBeat). With a context window of one million tokens it can accommodate an entire software project in one session.

The weights are being made available by Meta via Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. API access was in place from day one courtesy of Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The company has also put support for the leading serving stacks on the books, including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, complete with hardware optimizations for AMD, Apple, Intel and Nvidia.

It is a significant time to do so given that the open model scoreboard has swung hard in China's direction. According to Hugging Face's March review, Chinese models were responsible for 41 percent of downloads over the prior year, edging out those from the US. And in September OpenRouter said that open weight models were handling 60 percent of the token volume from US originating requests, the bulk of which came from Chinese labs.

Why Two Mac Studios Matter More Than the Benchmark Table

What drew the most notice was not a score. Meta had Muse Spark Open running on a pair of M5 Ultra Mac Studios from Apple, each boasting as much as 512 gigabytes of unified memory and 1.2 terabytes per second in bandwidth. By way of Apple's MLX framework the two are made to act as one system with about a terabyte of memory (Meta). In the company's view this is the first of its kind to be run on a desktop.

The desk setup is no bargain. The 256 gigabyte M5 Ultra with a one terabyte drive is put at $10,799 and the 512 gigabyte version is due in late October with no price yet set. As for the model card, it stipulates a minimum of 700 gigabytes of accelerator memory in NVFP4 precision and two terabytes in BF16, requirements that the 42 billion active parameters do little to abate.

The demos were to the point. According to Meta, with Muse Code operating entirely on the two desktops, iOS bugs were corrected, features added and full builds completed for Ice Cubes, the open source Mastodon client. DaVinci Resolve footage was gone over frame by frame to crop clips for vertical video. But no token rate or total hardware cost was put forward with those assertions.

"We're sharing this work because we want everyone to have access to the best models," Meta Superintelligence Labs wrote in the company's announcement.

Where Spark Open Wins, and Where Chinese Labs Still Lead

An honest tale is told in Meta's own comparison table. Spark Open has the edge over GLM 5.3 on JobBench, 61.8 to 61.4, and puts on a clear lead on MCP-Atlas at 89.2 versus 84.2. It also posts 26.1 on HLE-Diamond without tools, Kimi K3's number being 22.2. With tools for visual work it is 50 on ZeroBench to Kimi K3's 41.

In some of the agentic and coding tests the margins are reversed. GLM 5.3 comes in at 1,653 on GDPVal-AA v2, Spark Open at 1,401. On Terminal-Bench 2.1 Kimi K3 is 85.0 against 76.4 and has the upper hand on DeepSWE 1.1 with 68 to 60. As for the Artificial Analysis Intelligence Index, Meta's table has Spark Open at 39 and GLM 5.3 at 45; the independent leaderboard has MiMo-V2.6-Pro at 46 in first place. A separate public assessment of Spark Open from Artificial Analysis could not be verified when the story was published.

Those watching open models took the news with more relief than any sense of victory. "I waited so long for US open weight models to be relevant again," wrote Nathan Lambert, author of the Interconnects newsletter, in a post on X, adding that there is still more work to do.

Read the License Before You Build a Business on It

Rather than a conventional open source license, Muse Spark Open is put out under the Muse Model License. While it allows for the model to be run, modified, fine tuned, distilled and redistributed, there are conditions for any model as a service business whose revenue with affiliates passes $20 million in any twelve consecutive months: they have to put Meta on notice and enter into a written agreement of their own before making Spark Open or its offshoots available. Exemptions apply to internal use, research and products not in the model serving business, and products must carry a "Built with Muse Spark Open" notice. Meta does not assert ownership over what the model puts out.

Getting in on Hugging Face is more restricted than one might expect. The application calls for an applicant's full legal name and date of birth as well as the complete name of the organization; the form cannot be edited once submitted and the data is covered by Meta's privacy policy. On the commercial side the pricing is competitive. Together AI has it at $1.25 per million input tokens, $4.25 for output and $0.15 for cached input. Groq will only go up to 131,072 tokens of context whereas Fireworks can handle the full 1,048,576.

The initial partners have been concentrating on agent behavior. "This model performs particularly well on tool calling and long context, including multi-step and parallel calls with clean structured outputs," said Dong Chen, senior AI engineer at Ollama, in Together AI's launch post. "In our agentic engineering workflows, Muse Spark Open consistently selected the right tool and kept its responses compact, exactly the decisiveness production agents need," said Zimu Li, member of technical staff at Factory, in the same post.

From Llama to Muse: Meta's Second Open Weights Bet

This is not new territory for Meta. The first LLaMA was put out in February 2023 with as many as 65 billion parameters and a noncommercial license; by July of that year Llama 2 had been made available for commercial purposes. Llama 3.1's 405 billion parameter model came in July 2024 and, according to VentureBeat, Llama 4 added the multimodal Scout and Maverick models in April 2025.

Then there is the Muse line from Meta's reconstituted Superintelligence Labs. That unit rolled out the original Muse Spark on April 8 and on August 10 the Muse Glimmer, an open 30 billion parameter model under Apache 2.0 which can be compressed to less than 20 gigabytes.

An open release is not all there is, however. Meta has a closed tier as well. On October 12 it brought forward Muse Wonder and Wonder Full for professional and software engineering applications. Artificial Analysis has them at 52 and 56 respectively on its Intelligence Index (VentureBeat). The price for Wonder is $4 per million input tokens and $20 for output. Wonder Full is restricted to Muse Code subscribers at $50 or $200 a month; those plans come with a Contributor mode by default so Meta can train on what is submitted, unless the user elects otherwise.

As for Spark Open, Meta says it underwent further safety post training since open weights carry their own considerations. It has passed the company's Superintelligence Scaling Framework thresholds for release, though a complete evaluation report is yet to be issued (Meta).

What Happens When a Frontier Model Fits Under a Desk?

Economics of running agents would be the first to feel an impact. Should an 816 billion parameter model be able to run on a pair of desktops, firms with sensitive patient data, legal files or code may find it more prudent and cost effective to make a $40,000 hardware outlay than put every token through a cloud API. That is good news for Apple, not yet a major AI infrastructure supplier, but a silent menace to the hosted inference providers whose margins are a function of volume.

Founders will have to do some rethinking. An open US model that is up to the task and has permissive terms for those under $20 million in revenue could allow an early stage startup to put together agent products without incurring a per token charge and hasten their experimentation. On the other hand, the way Meta defines derivatives and its revenue clause means any company that becomes a model serving operation may have to come to terms with Meta; investors may begin to enquire about license exposure as they used to with cloud spend.

Competition may not thin out. Chinese labs are still ahead on a number of coding and agent benchmarks and if GLM, Kimi and Qwen continue to put out new versions every few months, Meta's lead on certain tests could be gone in a hurry. What to look for next: an independent score from Artificial Analysis for Spark Open, a price on the 512 gigabyte Mac Studio that puts a two machine configuration within reach of small teams, and whether Meta's data gate will drive developers to an Apache licensed option such as Muse Glimmer. Then there is the matter of the desktop demo. If it stands up to outside testing, in a year or two the serious work of local AI may move from hobbyist rigs to the average office.

What Investors Should Decode in Meta's Open Weights Play

Sparknify will be on hand to see what the Bay Area's founders and investors make of Spark Open, from the first startups to standardize on it to those who come up against the license. The question of open versus closed is not an academic exercise any longer, it is a line item on the pitch deck. What comes of the next year will tell if America's open weights revival is a one off or something more lasting, and that will determine who gets to build the agents we all use.

Subscribe to Sparknify for Silicon Valley tech news and the people behind it.