Muse Spark Open: Meta's 816B Open-Weight Model — Specs, Benchmarks, License, and Hardware Requirements

Meta released the weights for Muse Spark Open on October 15, 2026, describing it as "the most capable US open-weight model to date." The model is an 816B-parameter natively multimodal system with 42B active parameters and a 1M-token context window, positioned for coding, agentic work, and long-horizon reasoning tasks. The release drew 827 points and 415 comments on Hacker News, where the dominant themes were benchmark competitiveness against Chinese open-weight models, the terms of the license, and the cost of the hardware Meta used to demonstrate on-device inference.

What Muse Spark Open actually is

Muse Spark Open is a natively multimodal, open-weight language model with 816 billion total parameters and 42 billion active parameters, released under the Muse Model License. Meta states it is built for coding and agentic work, and that it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.

Attribute	Value
Total parameters	816B
Active parameters	42B
Context window	1M tokens
Modality	Natively multimodal
Release date	October 15, 2026
License	Muse Model License (limited commercial terms)
Primary positioning	Coding, agentic work, long-horizon tasks

Meta says the model is the first in its class to run on a desktop, using two Apple M5 Ultra Mac Studios with the MLX framework.

The desktop inference claim depends on two M5 Ultra Mac Studios

Meta's headline desktop claim requires two Apple M5 Ultra Mac Studios, not one. Meta's stated reasoning is that Apple's unified memory architecture lets the M5 Ultra's integrated GPU use up to 512 GB per Mac Studio, so two machines can hold the entire model, and the chip's 1.2 TB/s of memory bandwidth is fast enough to read the small slice each token needs. Meta describes the result as "two Mac Studios behave like one supercomputer with a terabyte of unified memory."

Hacker News commenters pushed back on the accessibility framing. One commenter calculated the cost:

"At the current market price of Apple Mac Studio Ultra, 30-core CPU + 64-core GPU w/ 256GB RAM is already reaching at around $10k. Keep in mind that computers are appreciating assets recently. The 512GB RAM option is not even for sales yet (they said late October availability). Based on typical Apple pricing, it's likely to be around $14k apiece. So we are looking at least $28k - $30k for the setup Muse team is advertising." — @devy

The same commenter compared the setup to an NVIDIA H100 with 96 GB of HBM3 and up to 3.94 TB/s of memory bandwidth, which is more than three times the Mac Studio's unified memory speed, and questioned how Meta obtained access to unreleased M5 Ultra hardware with 512 GB of RAM.

Several commenters treated the gap between Meta's "the future is for everyone" messaging and the hardware requirement as the release's central irony:

"'We believe AI is for everyone' *who can afford two M5 Studios with a total of 1 terabyte of memory." — @markstos

"Finally democratizing AI access to, checks notes, people with no less than two Apple M5 Ultra Studios on their desktop." — @octomind

A counterpoint in the thread argued that the Mac optimization is the genuinely important part of the release: "The exciting thing about this is in them optimizing for inference on Macs. This, and optimization for other consumer hardware, is where local AI needs to go" (@TomGarden).

Third-party benchmarks place it behind Chinese open-weight models

Meta's own benchmark table compares Muse Spark Open against Inkling, Nemotron 3 Ultra, Beam, Mistral 4 Large, Qwen3.8-2.4T-A95B, GLM 5.3, and Kimi K3 across agentic, knowledge and reasoning, coding, and tool-assisted multimodal evaluations. Meta points to a separate evaluation methodology report for details.

Commenters cited Artificial Analysis scores that place the model below several competitors. One commenter reported a score of 39 on Artificial Analysis, "about the same as Deepseek v4.1 flash and Mimo 2.5 Flash. Behind Glm 5.3 flash and Qwen3.8-Flash-Next" (@GodelNumbering). Another wrote that the model is "underperforming Qwen3.8 Flash Next and GLM 5.3 Flash, both of which you can run on a single B300" (@a11r).

A recurring complaint in the thread was that Artificial Analysis only publishes comparisons at the highest reasoning level, which commenters said is not representative of production use. "I would really like them to start providing comparisons at low and medium reasoning levels across these models because those tend to be the ones that actually get used in production," @a11r wrote.

Commenters also compared Muse Spark Open unfavorably to Meta's own prior models. "In terms of performance it seems like this has somewhat worse benchmarks than Muse Spark 1.3, closer to 1.2 but not the same. Of course significantly worse than Muse Wonder Full," wrote @modeless. Another commenter claimed the model "has an outdated attention mechanism which at 1M tokens will need over 200GB just for the KV cache" (@redox99).

One commenter questioned the framing of the benchmark set itself: "Its interesting that muse benchmarks are limited to open-weight models. while when glm introduced 5.3 and its flash variant they were compared to kimi, Anthropic and Open ai" (@Regex777).

The license is open weight, not open source

The Muse Model License allows free use for many developers but includes a Model-as-a-Service clause. One commenter read the license and summarized the key restriction:

"If the Licensee or any of its affiliates provide a third party with access to artificial intelligence models as a product or service, via API or a similar interface, where the third party receives outputs of such models ('Model as a Service'), and the aggregate gross revenue of the Licensee and its affiliates exceeds twenty (20) million US dollars ... during any consecutive twelve (12) month period, the Licensee must notify Meta and enter into a separate written agreement with Meta before offering the Software or any Derivatives as part of that Model as a Service business." — @seizethecheese, quoting the license

The same commenter noted the clause does not apply to internal use, research, or commercial activity that is not Model as a Service, nor to using the model as a component inside a product that is not itself a model service. "They're allowing the use of this in commercial products, just not selling the model directly," the commenter concluded.

Other commenters objected to the framing. "It's not open!!!" wrote @I_am_tiberius. A more measured take: "This is great progress for 'open-weight' models. Quotes due to the restrictions, but still progress none-the-less" (@itsmeduncan).

One commenter argued the separate-licensing requirement has a market effect: "The usual effect of separate licensing is decreased price competition from public providers (like we saw with Kimi K3). I guess it's better for the providers if the price is fixed: any efficiency improvements directly improve margin rather than driving prices down over time" (@popinman322).

Inference support spans nearly every major runtime

Meta states Muse Spark Open runs on SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with optimizations for AMD, Apple, Intel, and NVIDIA silicon. Developers can further customize the model using PyTorch, and Meta points to TorchTitan for tuning.

Commenters noted the breadth of this support as a genuine strength: "They've done considerable effort in making sure this model can run on basically every model runner, from Unsloth to MLX" (@armcat).

The open-weight release ships with custom, auto-tunable Helion GPU kernels in SGLang. Meta reports 1.03–3.95× kernel-level speedups on NVIDIA and AMD GPUs and a 3% end-to-end speedup on AMD MI350.

Availability: APIs, weight hosts, and agentic demos

Muse Spark Open is available through API platforms at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Model weights can be downloaded from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. Meta published an MLX adapter for running the model on Mac Studio.

Meta reported testing agentic workflows running entirely on-device using Muse Code powered by Muse Spark Open, including shipping real iOS changes to Ice Cubes, an open-source Mastodon client — bug fixes, new features, and a full build — and driving DaVinci Resolve end to end, with the model reviewing footage frame by frame to find the strongest moments and cropping them vertically.

One commenter reported running image-to-HTML conversion tests and found the model strong for that task: "Overall it actually does pretty solid. It doesn't completely follow the design, but it gets the vibe. It's one of the best open models I've seen for conversion, which is impressive" (@jjcm).

Safety evaluation and licensing framework

Meta states that Muse Spark Open received additional safety post-training specifically because open weights carry their own considerations. The company says it evaluated the model in line with its Meta Superintelligence Scaling Framework, covering the most severe risks, and that the model met the thresholds the framework requires for release. Detailed results are in the Muse Spark Open model card on Hugging Face, with a full evaluation report promised.

The strategic argument for releasing the weights

The thread contained a substantial debate about why Meta would release a frontier-adjacent model at all. The most-upvoted strategic take argued that Meta's original open-model thesis is being validated:

"The rent market on intelligence never materialized, if you have THE frontier model, you can charge rent (for a few weeks, maybe month). But after a certain capability level it's price that drives the decision. Zuck is doing the thing he's really good at, and building a moat around the network (Muse is VERY sticky), and now that we're all going to standardize on Meta's model, we're going to optimize the heck out of the infra to run the model making it cheaper for him to scale, thus bringing in more participants to his network, and thus making his network more valuable. Meta is going to win consumer AI." — @swalsh

A related argument focused on distribution rather than model quality: "there are not that many teams in the world that can deploy muse like product to A/B test a lot of hypothesis at the scale meta can at scale ... My mom and dad and their family have never heard of Claude and don't use chat gpt except once in a while as a search engine. But meta will figure out a way to insert this into modals and feeds" (@itissid).

Others were skeptical of the "US open-weight" framing itself. "What is the point of releasing 'The Most Capable US Open‑Weight Model?' I can understand why I might care about the location of a model serving provider, but why should I as a model user care about the location that an open weights model was trained?" asked @lern_too_spel. Another commenter observed a pattern across labs: "companies doing open weight models are shifting from advertising them as 'the best' to 'the best in the US/EU' ... it also sounds like an implicit acknowledgment of how good Chinese companies have become in the space of open-weight models" (@sergdigon).

Cost and pricing concerns for API users

Commenters flagged pricing as a practical blocker. "Prices need to come down on openrouter because right now it's a model that's not as good as deepseek v4.1 flash and it costs about 10x" (@keyle). Another commenter noted that the main thing keeping them on a paid subscription is token costs for the more capable models (@popinman322).

Open questions raised by the release

Several points in the thread remained unresolved by the source material:

• Whether Meta had access to unreleased M5 Ultra hardware with 512 GB of unified memory, or whether the desktop demonstration was written from specifications. A commenter noted the 512 GB configuration was listed for late-October availability (@devy).
• How Muse Spark Open compares to Muse Spark 1.3 and Muse Spark 1.2 on identical evaluations, given commenters' claims that the new release benchmarks below both.
• Whether the 1M-token context window is practically usable given the KV cache memory requirements one commenter estimated at over 200 GB.
• Whether Meta will release weights for Muse Spark 1.2, which one commenter said Zuckerberg had personally committed to (@sharktheone).

Meta's full evaluation report was described as "available soon" at the time of release.