Generated by Codex with GPT 5.6 Sol High

Techmeme surfaced Meta’s October 15 release, “Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date”. The important part is less the superlative than the deployment target: Meta has released a large multimodal agent model whose compressed weights can fit across two high-memory Mac Studios. That moves a class of capability usually associated with a datacenter into hardware that a well-funded lab, company, or individual can own and operate without sending prompts to a model provider.

A sparse giant that can fit on a desk

Muse Spark Open is a mixture-of-experts model with 816 billion total parameters, but it activates only 42 billion for each token. The inactive experts still have to be stored, so sparsity reduces the computation needed for inference more than it reduces the memory needed to hold the model. Meta’s model card says the NVFP4 version requires at least 700 GB of GPU memory, while the BF16 version requires 2 TB.

That is where Apple’s memory architecture matters. Each M5 Ultra Mac Studio can make as much as 512 GB of unified memory available to its integrated GPU. Linked together, two machines provide roughly a terabyte of addressable memory, enough for the quantized checkpoint. Meta says MLX turns the pair into one inference system and that the M5 Ultra’s 1.2 TB/s memory bandwidth can feed the small subset of weights selected for each token.

Calling this “desktop” AI is therefore both fair and easy to misread. The model is not running on an ordinary laptop or a single consumer GPU. It needs two unusually large-memory workstations, or a server with hundreds of gigabytes of accelerator memory. But it is still a meaningful change in kind: an organization can keep the weights, prompts, files, and tool traces on equipment it controls, continue operating without an API vendor, and customize the model locally.

The model accepts text, images, video, and audio and produces text, with a one-million-token context window. Meta positions it for long-running coding and tool-use workflows. Its demonstrations included making and building changes to the open-source Ice Cubes iOS client and operating DaVinci Resolve while examining video frame by frame. These are demonstrations chosen by the developer, not independent measures of reliability, but they show why the combination of multimodality, long context, and local execution is more consequential than a chatbot benchmark alone.

Strong agent scores, not a universal winner

Meta’s published numbers support a narrower claim than “best open model.” Muse Spark Open scored 89.2 on MCP-Atlas, ahead of the listed competitors, and 61.8 on JobBench, approximately level with GLM 5.3’s 61.4. It also reached 80.4 on OSWorld Verified, behind Kimi K3’s reported 84.8 but well above Meta’s earlier Inkling model.

The broader table is mixed. Muse scored 39 on the Artificial Analysis Intelligence Index, below Qwen 3.8 at 40, Kimi K3 at 44, and GLM 5.3 at 45. Its coding results also trailed several peers: 60 on DeepSWE 1.1 versus 67 for GLM 5.3 and 68 for Kimi K3, and 76.4 on Terminal-Bench 2.1 versus Kimi K3’s 85.0. The release looks especially competitive in tool use and multimodal work, not uniformly dominant across reasoning and coding.

Those comparisons also mix several evidence types. Meta used official leaderboards where available, internal runs in some cases, and model-makers’ self-reported results in others. Coding models sometimes ran through different “preferred” agent harnesses, which can materially affect results. Meta also notes that its common setup was not tuned for every competing model. The table is useful evidence that Muse Spark Open belongs among leading open-weight systems, but it is not a controlled tournament proving one overall ranking.

Open weights with durable tradeoffs

Meta is releasing the parameters rather than the complete recipe needed to reproduce training, so “open-weight” is the precise description. The Muse Model License permits use, modification, fine-tuning, distillation, redistribution, and sale, but it is not an unrestricted open-source license. Products must display “Built with Muse Spark Open,” uses must follow Meta’s policy, and companies with more than \$20 million in annual revenue that offer models as a service must reach a separate agreement with Meta before including Muse or its derivatives.

The safety arrangement reflects the same compromise. Meta added extra post-training to the downloadable checkpoint and reports very low attack success on its StrongREJECT evaluation, high refusal rates on cyber and biological misuse tests, and no misbehavior on one agentic-misalignment benchmark. Yet a holder of the weights can fine-tune or otherwise modify the model after release. Meta explicitly warns that its safety findings apply only to the tested checkpoint and recommends independent guardrails and human confirmation before irreversible agent actions.

That limitation cannot be solved by a terms-of-use document alone. Local weights improve privacy, auditability, customization, resilience, and competition because the provider no longer mediates every interaction. The same separation weakens the provider’s ability to monitor abuse, revoke access, or push a safety update. Muse Spark Open makes both sides of that bargain more concrete because it combines downloadable weights with strong agent and multimodal capabilities.

The release’s lasting significance is therefore architectural. Sparse models, aggressive quantization, high-bandwidth unified memory, and portable inference frameworks are shrinking the boundary between “local” and “datacenter” AI. Two Mac Studios are not mass-market hardware, and Muse is not the best model on every benchmark. But a capable, million-token, multimodal agent can now be placed under its operator’s physical and administrative control. That is a real expansion of who can run frontier-adjacent systems—and of who must manage their consequences.