Deep Dive — America's top open model still trails China's best

Meta released the weights for Muse Spark Open on Thursday and called it "the most capable US open-weight model to date." Independent measurement agrees with the first half of that sentence and complicates the second: on Artificial Analysis's Intelligence Index the model scores 39, first among American open weights and behind GLM-5.3 at 45, Kimi K3 at 44 and Alibaba's Qwen3.8 at 40. The qualifier in Meta's headline is doing the work. This is a real release, a real engineering achievement, and not a rebalancing of the open-model field.

What actually shipped

Muse Spark Open is a mixture-of-experts model with 816 billion total parameters and 42 billion active per token, a 1-million-token context window, and text output. Meta's announcement describes input across text, images, video and audio; Artificial Analysis lists its input modalities as text and image only, a gap worth watching as the model card matures. Weights are on Hugging Face in BF16 and NVFP4.

The architecture is aggressive on sparsity. Per vLLM's day-zero implementation notes, the model stacks 62 decoder layers, all of them MoE, with 256 experts per layer and 8 active per token — a 1-in-32 activation ratio, sparser than most of the trillion-class open models it is being compared against. A latent-MoE pooler projects the hidden state from 8,192 down to 4,096 for expert computation and back up, and attention is hybrid: 46 layers use a 2,048-token sliding window, 16 attend to the full context. Vision runs through a separate Perception Encoder, audio through a 118M encoder. Elie Bakouch notes it is the only major US open model without shared experts.

Meta's own benchmark table is strong where it can be, and Meta selected its columns. MCP-Atlas tool use: 89.2, ahead of Qwen3.8 at 84.5 and GLM-5.3 at 84.2. JobBench: 61.8, just past GLM-5.3's 61.4. HLE-Diamond without tools: 26.1 against Kimi K3's 22.2. ZeroBench with tools: 50.0 against Kimi K3's 41.0, and a top-of-table 25 on AA-Omniscience. Then coding, which is the weak spot: 76.4 on Terminal-Bench 2.1 with the Terminus 2 harness, behind Kimi K3 at 85.0, GLM-5.3 at 83.9 and Qwen3.8 at 82.0.

The qualifier is the story

"Most capable US open-weight model" is a claim about procurement, not about capability, and American buyers should read it that way. The independent index number matters more than the geography: 39 puts Spark Open behind three Chinese open models and, by VentureBeat's reading of the published numbers, behind Xiaomi's MiMo-V2.6-Pro at 46. Artificial Analysis's breakdown is blunter still — Spark Open ranks last among the nine highest-scoring open models on the index's four agentic evaluations, and GLM-5.3's entire six-point lead comes from that one segment.

That is the segment that matters most for the workload Meta built this model for. Long-horizon agents and multi-file coding are agentic evaluations, so the deficiency lands exactly where the marketing points. The counter-argument is real: Spark Open uses about 40,000 output tokens per index task, the fewest among those nine models, so it reaches its score without burning tokens on deliberation, and it does so at $1.25 per million input and $4.25 per million output tokens — launch pricing on Together, Fireworks and Baseten, $0.15 for cached input. Artificial Analysis calculates $0.84 per index task, against $2.01 for GLM-5.3 and $0.27 for DeepSeek V4.1 Flash. Cheap, concise, second-tier on agentic work: a defensible product for a lot of teams, and not the frontier.

Two data points explain why Washington cares about the geographic framing anyway. Hugging Face's March 2026 assessment found Chinese models accounted for 41 percent of downloads over the previous year, passing US models. OpenRouter reported in September that open-weight models handled 60 percent of token volume from US-origin requests on its platform, with Chinese labs supplying most of that traffic. The strategic worry is not that an American bank cannot find a good open model; it is that the default open model in most stacks is Chinese. Muse Spark Open narrows that gap — Nathan Lambert's read is that it cuts the US–China open-model gap roughly in half — without closing it.

Meta shipped the weaker model

The part that should unsettle the "open weights are a safety valve" argument is lineage. Jack Rae says Muse Spark Open was the first base model trained under Meta's new Superintelligence Labs organisation, with additional training applied for the open release. Several researchers reading the model card described it as roughly Muse Spark 1.2 level — and Meta's closed flagship is 1.3, which in September claimed the top long-horizon coding score in the field and which Artificial Analysis put at 62 on the same index at maximum reasoning effort. That is the pattern we flagged when the weights landed, in Meta opens 816B Muse Spark weights — and two Macs can run them: the open release is a derivative of a lineage the company has already surpassed. If you want to see how far the closed side has run, the September release is the reference — Meta's Muse Spark 1.3 tops DeepSWE — and undercuts the frontier.

This is now a repeated shape rather than an exception, and it is not unique to Meta. Alibaba's Max-class open release was a step behind its own frontier, and StepFun's open weights arrived with a bespoke licence and a two-node deployment. Labs have stopped treating their best work as the gift and started treating the previous generation as distribution.

The licence is now a template

Spark Open ships under the Muse Model License. It grants use, copying, modification, fine-tuning, distillation, merging, publishing, distribution and sale — and then attaches two conditions. Every product or service using the model or anything derived from it must prominently display "Built with Muse Spark Open" in its interface, documentation, website or settings menu. And any licensee that provides third parties with access to AI models via an API or similar interface — Model as a Service — must notify Meta and sign a separate written agreement once the licensee and its affiliates exceed $20 million in aggregate gross revenue over any consecutive twelve months. Internal use, research, and the use of the model as a component inside a licensee's own product are carved out.

Read that against the two licences that shipped in the days before it and the template is obvious. Alibaba's Qwen3.8-Max licence puts a separate-commercial-licence gate at $50 million in revenue in any twelve-month period, plus a naming requirement above 100 million monthly users or $20 million in monthly revenue. StepFun shipped Step 5 Preview under a licence it describes as Apache-2.0-style with additional attribution and commercial conditions. Hard gates, revenue thresholds and attribution clauses are becoming standard furniture on frontier-scale "open" releases, and the thresholds are calibrated to catch exactly the businesses that turn open weights into a competing API business. If you are the one making that call, our walkthrough on How to — check a model's license before you ship on it is the pre-flight checklist. The terms are negotiable, not fatal — but the day your revenue crosses the line, you are no longer shipping free software, you are in a commercial relationship you have not priced.

Trade coverage adds a restriction the excerpted licence text does not show in the section we read: SiliconAngle reports military and critical-infrastructure uses are off-limits. Treat that as reported rather than verified until the full text is parsed.

What "runs on your desktop" costs

The deployment story is genuinely new and genuinely expensive. BF16 weights come to roughly 1.6 TB. The NVFP4 checkpoint quantises only the experts and lands near 530 GB, which Meta says runs on four Blackwell Ultra GPUs, eight H200s, or two 512 GB M5 Ultra Mac Studios over MLX. Apple has not published pricing or a ship date for the 512 GB configuration beyond "late October." The two-machine headline has a cheaper cousin: four 256 GB M5 Ultra Mac Studios at roughly $38,000, per Engadget, or a single 256 GB M5 Ultra Mac Studio, which starts at $10,799, if you are willing to run a quantised build.

Those quants are what make the claim real for anyone who is not an enterprise. llama.cpp had GGUF support on day zero, and Unsloth's UD-IQ1_S build lands at 176 GB. A 3-bit MLX quant still needs more than 256 GB of memory, per Hugging Face's own guidance, so the practical floor for local use is one heavily quantised machine — and the practical ceiling on quality at 1-bit is a separate question nobody has answered publicly yet. The 1-million-token context is the same story in a different currency: a long-context agentic session on a 42B-active model does not run on consumer memory without either a cluster or aggressive quantisation, which is why the honest local configuration is two machines rather than one.

What the skeptics say

Three objections hold up. First, the index gap and the benchmark gap are measuring slightly different things, and Meta's table is self-reported where Artificial Analysis's number is not — so the fairest reading is that Spark Open is competitive on tool use and knowledge reliability, clearly behind on terminal coding and agentic evaluations, and the disagreement is about magnitude rather than direction. Second, the sovereignty pitch is weaker than it sounds: OpenRouter already offers US hosting for Chinese open models, so a US enterprise that needs its inference to happen on American soil can have that today with a DeepSeek or Kimi model. Where a model was trained is not where its data is processed. Third, the safety post-training Meta added for the open release has a cost that Reddit's technical thread picked up immediately — a commenter working in security argued the extra tuning makes the model a poor fit for cybersecurity workflows and recommended GLM-5.3 instead, which is a strange place for a US open release to lose to a Chinese one. The grim read, from commentator teortaxes, is that the result is thin at this parameter count with dated attention, and that Spark 1.3 is substantially stronger.

The fair criticism is not that Meta shipped a bad model. It is that "most capable US open-weight model" is a phrase engineered to be true while sounding like something larger, and that the useful comparison — does this change what you can run that you could not run last week — depends on quantisation quality, licence thresholds and a Mac configuration you cannot buy yet. The answer to that question is genuinely yes for a research team that needs inspectable frontier-scale weights, and no for anyone hoping the open side caught up.

What to watch

• Whether an independent host publishes sustained throughput and quality numbers at the full 1-million-token context rather than short-context smoke tests.
• Whether a credible third-party reproduction of the agentic evaluations lands anywhere near Meta's table, particularly on Terminal-Bench 2.1.
• What the 512 GB Mac Studio actually costs and when it ships — the local-inference story is priced off a machine Apple has not quoted.
• Whether Meta's next closed model opens the gap again, and whether the open release keeps trailing by exactly one generation.
• Whether the $20 million Model-as-a-Service threshold becomes the industry norm or gets undercut by a lab willing to ship without it.

Meta's open weights are the second-best American option by definition and the third-best globally by measurement. Does "most capable US" change your procurement, or is capability the only line that matters? Tell us in the comments.