Meta's Muse Spark Open Claims the US Open-Weight Crown, But the Title Comes With Fine Print

Meta has released Muse Spark Open, an 816-billion-parameter open-weight model it markets as "the most capable US open-weight model to date." The claim is narrower than it sounds, and enterprise buyers should read the qualifier closely before treating this as a performance leader.

Chinese labs have dominated open-weight adoption for over a year. Hugging Face's March 2026 assessment found Chinese models accounted for 41% of downloads over the preceding year, overtaking US models for the first time. OpenRouter reported in September that open-weight models handled 60% of token volume from US-originating requests on its platform, with Chinese labs supplying most of that traffic. Against that backdrop, a credible US entrant matters strategically, even if it doesn't top every leaderboard.

Meta's own model card reports an Artificial Analysis Intelligence Index score of 39 for Spark Open. That trails Xiaomi's MiMo-V2.6-Pro at 46, GLM-5.3 at 45, Kimi K3 at 44, GLM-5.3-Flash at 42, and both Qwen3.8 2.4T A95B and Qwen3.8-Flash-Next at 40, according to Artificial Analysis's independent listings. All six of those published scores exceed Spark Open's. Meta's own benchmark table does show Spark Open ahead of listed rivals on two specific tests, JobBench (61.8 versus GLM 5.3's 61.4) and MCP-Atlas (89.2 versus 84.2 for GLM 5.3). But Kimi K3 and GLM 5.3 lead on terminal coding, extended software-engineering tasks, and the GDPVal-AA professional-work evaluation.

That split matters for procurement teams. Spark Open is a contender with specific strengths, not an across-the-board leader. Enterprises weighing task completion, reliability, licensing terms and operating cost will find Meta's case rests on fit for particular workloads rather than universal superiority.

Hardware Promise, Hardware Catch

Meta's headline pitch, and arguably its most attention-grabbing one, is that Spark Open can run locally on a pair of M5 Ultra Mac Studios. Each machine offers up to 512GB of unified memory; paired, Meta says, they provide a terabyte and MLX enables local operation. The company demonstrated Muse Code, powered by Spark Open, modifying the Ice Cubes Mastodon app and operating DaVinci Resolve to review footage and prepare vertical crops.

The catch: Apple's 512GB-memory configuration doesn't ship until late October, and its price is unannounced. Current pricing offers only a rough proxy. Apple lists a 256GB M5 Ultra workstation with a 1TB SSD at $10,799; four units would match the combined terabyte of memory for $43,196 before tax and accessories. That's a reference point for assembling comparable capacity today, not a confirmed price for Meta's two-computer setup.

The underlying memory math deserves scrutiny too. Spark Open's model card lists at least 700GB of accelerator memory for the NVFP4 version and 2TB for BF16. The model activates only 42 billion of its 816 billion total parameters per token through a mixture-of-experts design, but that efficiency doesn't shrink the stated memory footprint down to something resembling a standalone 42-billion-parameter model. Meta's announcement also supplies no measured token-generation rate for the desktop setup and no complete hardware cost. These are company demonstrations, not independently verified benchmarks.

On licensing, the Muse Model License grants rights to run, modify, fine-tune, distill and redistribute the model. Its revenue trigger is specific: when a licensee and its affiliates generate more than $20 million in aggregate gross revenue over any consecutive 12 months, the license requires notification to Meta and a separate written agreement before offering Spark Open within a model-serving business. That threshold covers total company revenue, not revenue tied specifically to Spark Open, but it explicitly exempts internal deployment, research, and use inside a company's own product that isn't itself a model-as-a-service offering. For most enterprises building an internal assistant or embedding AI into an application, that exemption removes the $20 million figure as a practical ceiling.

Access has its own wrinkle. Spark Open's Hugging Face gate asks applicants for full legal name, date of birth, and full organization name with corporate identifiers, warning that non-compliance could affect access to other models on the platform too. That's more detailed than Hugging Face's default gate, which typically just shares a username and email with model authors. The practice has precedent: Meta's Llama 4 Scout, Meta-SecAlign-8B and OMol25 carry identical notices, as does Trillion Labs' Tri-21B. It establishes a pattern, not a universal standard, and enterprise evaluators should separate the question of what they're permitted to do with downloaded weights from what personal data a given repository demands to hand them over.

On pricing, Together AI lists Spark Open API access at $1.25 per million input tokens, $0.15 per million cached input tokens, and $4.25 per million output tokens, for a blended $5.50 per million tokens. That sits above several Chinese competitors on a straight cost basis, including MiMo-V2.6-Pro at $1.305 per million blended and DeepSeek-V4-Pro's off-peak rate of $2.64. Meta's announcement does not establish a cost lead; it omits hosted token rates in its own materials, leaving third-party hosts like Together AI, Fireworks, Groq and Nebius to set the terms buyers actually see. Deployment details vary by host too: Groq's documentation caps context at 131,072 tokens and requires organizational approval, while Fireworks lists support for the full 1,048,576-token context Meta advertises.

The Bottom Line

Muse Spark Open is a legitimate addition to the US open-weight roster, not a definitive leader. Its Intelligence Index score of 39 trails six independently measured Chinese models, even as it beats specific US rivals like Inkling (26) and Nemotron 3 Ultra (23). The desktop story depends on Mac Studio hardware that isn't shipping yet at an unannounced price, and the licensing and token economics look competitive rather than disruptive. For enterprise buyers with no geographic procurement constraint, Spark Open earns a seat at the table alongside GLM 5.3, Kimi K3 and Qwen3.8, not a seat at the head of it.