Start with the two documents, because the models are almost beside the point. One is a bespoke Muse Model License demanding that any product or derivative prominently display "Built with Muse Spark Open," and requiring any Model-as-a-Service business above $20M in aggregate annual revenue — the Muse Model License requires Model-as-a-Service businesses with more than $20M in annual revenue to sign a separate agreement with Meta; Interconnects describes the threshold as $20M per month. The other is an unmodified Apache 2.0 header. They shipped one day apart (October 15 and October 16, 2026).
That juxtaposition is the news. The question is no longer whether a US lab can build a frontier-scale open-weights model. It is what "open" is permitted to mean now that two of them did.
The scoreboard matters first. Meta released Muse Spark Open on October 15: 816B total parameters, 42B active, a 1M-token context, native text, image, video, and audio input. Reflection released Beam's weights on October 16: 501.3B total, 23.4B active, 1M context, text in and text out. Interconnects called it a step change back toward the frontier — and noted the counterfactual that gives the week its weight. Without these two models, there would be 20-plus Chinese-built open-weight models better than the best Western ones.
Without these two models, there would be 20-plus Chinese-built open-weight models better than the best Western ones.
The gap roughly halved, according to Interconnects.
Now read the licenses side by side, because "open-weight" here means two incompatible things. Meta's license frees the weight but binds the behavior: the attribution byline travels into products and derivatives, the $20M trigger is measured against the licensee's and its affiliates' total revenue rather than revenue earned from the model, and an acceptable-use policy rides along into outputs and derivatives — creating legal ambiguity about whether you can safely use the model for synthetic data. Beam's Apache 2.0 grant carries none of that. The repositories are public and ungated, and the model card explicitly contemplates commercial use, fine-tuning, quantization, merging, and distillation.
Nathan Lambert's verdict, relayed through the roundups, was blunt: the Muse license is "trash." That word is doing real analytical work. It is not a claim about model quality — Muse Spark Open scored 39 on Artificial Analysis's Intelligence Index, the highest for a US open-weights model, 13 points above NVIDIA's Nemotron 3.5 Super VL, and it leads open models on AA-Omniscience. It is a claim about the artifact's legal shape.
Here is where the two bets diverge. Meta is treating openness as a distribution strategy: give away the weights, keep a claim on the surface where products are built, and convert the largest commercial users into counterparties. It is the Llama playbook with sharper teeth. Reflection is treating openness as the product: a plain permissive grant, three checkpoint formats, a speculative-decoding drafter, and an open Rust-based harness, Mirror, that Beam was trained alongside, plus $10,000 in Hub credits to whoever builds the best thing with it.
The hardware undercuts the easy version of this story, and this is the piece most readers get wrong. Meta's "runs on your desktop" claim rests on two 512GB M5 Ultra Mac Studios — and Apple hasn't shipped the 512GB configuration yet. The 256GB M5 Ultra workstation lists at $10,799; four of them would approximate the same terabyte of memory for $43,196 before tax. The model card lists at least 700GB of accelerator memory for the NVFP4 version and 2TB for BF16. This is a five-figure desk, not a laptop download.
Beam is not meaningfully cheaper to own. The BF16 checkpoint is about 1TB, Hugging Face recommends four B200s for practical NVFP4 serving, and the documented BF16 vLLM recipe spans 16 H200s. Its 1M context is a configuration ceiling rather than a retrieval-accuracy result; the key-value cache alone grows about 36 KiB per token per sequence.
So the accessibility gap is real but roughly symmetric. Both artifacts are infrastructure projects. Neither is a download-and-run model, which means the licensing difference is not being mediated by who can actually hold the weights — it is being mediated by everyone else.
That reframes the restrictive license as less enforceable than it reads. Cheap hosted inference is the pressure valve: per Artificial Analysis, Muse Spark Open runs at roughly $0.84 per index task, against $0.27 for DeepSeek V4.1 Flash. When the inference layer is where most builders meet the model, a byline requirement and a revenue conversation with Meta become friction at the edges rather than gates at the center. A Hacker News commenter made the commercial version of this point precisely: "The usual effect of separate licensing is decreased price competition from public providers."
Which is the uncomfortable synthesis. The freedom that matters now is commercial, not technical — and the community's instinct to ignore licenses it finds trashy is exactly why licenses are the real strategic battleground. Interconnects found that Western open-weight models get astonishingly more adoption than Chinese models with the same performance-to-size footprint, even though the Chinese models usually ship first. Restrictive licenses have not killed adoption yet. But they remain, in Interconnects' phrasing, an "unnecessary strategic risk" for any lab trying to compete for that adoption.
The stakes are the post-training economy and the buyers behind it. Enterprises want to own their intelligence; governments want sovereign deployment. Reflection's partner list — 40-plus integrations spanning managed endpoints, cloud platforms, on-premises systems, and sovereign infrastructure — reads as a direct bid for exactly those buyers, with the license cleared as a procurement question before the first meeting. Meta's terms make that same conversation longer and more expensive.
Both models land into an environment where open-ended LLM agents are finally real and open-model demand is accelerating. That is precisely when license terms stop being a debate among researchers and start being a line item in a legal review.
There is a defensible reading in which Meta's license is simply the price of the distribution: it wants attribution and a cut of the largest resellers, and it will enforce neither against the long tail. The brand travels; the small builder stays unbothered. But the asymmetry is not free. Beam's Apache 2.0 grant lets any provider, fine-tuner, or synthetic-data pipeline build without a legal question, and lets any government buyer sign without an exception request. Meta's grant requires a conversation with Meta above a revenue line that is measured against the whole business.
The weights are comparable. The permissions are not. And in a market where model performance clusters within a few index points, the permission is the differentiator that scales.
Muse Spark Open is the smarter model by the published index — 39 against Beam's 30. Beam is the more open artifact by any conventional definition. That inversion is the week's real lesson, and it will not resolve on a leaderboard. It will resolve in procurement offices, where "Built with Muse Spark Open" either appears in a subtitle or it doesn't, and where the question is not how good the model is but who is allowed to sell it.