On October 15, 2026, Meta put an 816-billion-parameter model on Hugging Face and called it the most capable open-weight model to come out of the United States. The Hacker News thread reached 920 points and 478 comments, and a companion Apple machine-learning post about running the thing on Mac Studios pulled another 363 points on its own. Muse Spark Open is genuinely downloadable, genuinely free for most people, and genuinely very large — 816B total parameters with 42B active, a one-million-token context window, and native multimodal input across text, image, video and audio. But two details decide whether "open" means anything to you in particular: the floor on hardware, which starts around 700GB of VRAM, and a single clause in the license that reaches past you and sets the price of every API you might rent it through instead.
What Meta Actually Shipped
The architecture is a mixture of experts: 256 experts total, 8 firing per token, which is how a model with 816B parameters on disk only pays for 42B of them on each forward pass. A 1.8B-parameter vision encoder ships alongside it. Weights are posted on Hugging Face and mirrored into Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry's model catalog, and the model runs on SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth. That is an unusually complete day-one surface area, and it is the strongest thing about this release.
The Safety Post-Training Is the Interesting Technical Claim
Meta's research post says the model got additional safety post-training specifically because it is open-weight — the reasoning being that a hosted model sits behind system-level guardrails the provider controls, while a downloaded model has none, so the behavior has to live in the weights themselves. Meta evaluated it against what it calls the Meta Superintelligence Scaling Framework for severe risks and says it met the release thresholds. The published numbers back the direction: StrongREJECT v2 at 0.5 (lower is better) and FORTRESS at 5.8 harmful-answer rate. The cost shows up in the over-refusal column, 17.4 on FORTRESS ORS — meaning roughly one in six borderline-but-benign prompts gets declined. If you are building on this, that is the number to test against your own traffic before anything else.
The Scores Are Good, Not Leading
Here is where Meta's "most capable" framing needs its qualifier read out loud. The claim is specifically about US open-weight models, and the Artificial Analysis Intelligence Index does not put Muse Spark Open on top of the open field overall:
| Model | AA Intelligence Index | DeepSWE 1.1 | Origin |
|---|---|---|---|
| GLM 5.3 | 45 | 67 | China (Z.ai) |
| Kimi K3 | 44 | 68 | China (Moonshot) |
| Qwen 3.8 | 40 | 57 | China (Alibaba) |
| Muse Spark Open | 39 | 60 | US (Meta) |
| Mistral Large 4 | 38 | not reported in sources | France (Mistral) |
Muse Spark Open's own strongest results are elsewhere: 89.2 on MCP-Atlas, 76.4 on Terminal-Bench 2.1, 61.8 on JobBench. Read together with the 60 on DeepSWE, that shape says tool-calling and terminal-driving agent work rather than puzzle-style coding — the same profile Meta's researchers describe when they talk about long-horizon tasks and multi-step reasoning. One more caveat worth carrying: Hacker News user a11r pointed out that Artificial Analysis publishes comparisons only at the highest reasoning level, which is close to useless for high-volume production, and RussianCow added that open-weight models rarely get the lower-effort scores proprietary models publish. Muse Spark Open's reasoning effort defaults to xhigh — the most expensive setting — so the benchmark configuration and the sane production configuration are not the same configuration.
The Hardware Floor Is the First Catch
The download is not the hard part. The NVFP4 quantization lands at roughly 493–530GB; full BF16 is about 1.6TB. The hard part is what has to hold it in memory at once.
The GPU Path
Minimum practical VRAM is around 700GB for the NVFP4 build and about 2TB for BF16. The recommended configurations are four Blackwell Ultra GPUs or eight H200s. There is no consumer-GPU story here at all — not a 5090, not two of them, not a quantization trick that gets you under the line. For contrast, a11r noted in the thread that both Qwen 3.8 Flash Next and GLM 5.3 Flash run on a single B300, and colordrops and jnaina reported Qwen 3.8 Flash Next running on a 5090 with 96GB of system RAM, describing it as fast and good. That is the comparison that stings: the models beating Muse Spark Open on the index are also the ones that fit on hardware a person can own.
The Mac Studio Path, Which Is the One People Actually Tried
Meta worked with Apple on MLX optimizations and on Thunderbolt 5 RDMA for distributed inference, and the result is the most-discussed deployment of the launch. A single M5 Ultra Mac Studio cannot fit the model. Two of them, 512GB of unified memory each, produce somewhere between 15.74 and 24.99 tokens per second depending on the measurement. Four get you 36.73. At $5,499 per machine, the entry ticket is $10,998 for about a terabyte of memory and output slower than most people type-read comfortably.
| Deployment | Hardware | Approx. cost | Throughput | Context available |
|---|---|---|---|---|
| Two Mac Studios (M5 Ultra) | 2 × 512GB unified memory | $10,998 | 15.74–24.99 tok/s | Up to 1M |
| Four Mac Studios (M5 Ultra) | 4 × 512GB unified memory | $21,996 | 36.73 tok/s | Up to 1M |
| Datacenter GPU, NVFP4 | 4 × Blackwell Ultra or 8 × H200, ~700GB VRAM | not disclosed in sources | not disclosed in sources | Up to 1M |
| Datacenter GPU, BF16 | ~2TB VRAM | not disclosed in sources | not disclosed in sources | Up to 1M |
| Hosted API (Baseten, Fireworks, Together, Modal, Nebius) | None | $1.25 in / $4.25 out per 1M tokens | not disclosed in sources | Up to 1M |
| Groq (limited preview) | None | not disclosed in sources | 350+ tok/s | 131,072 |
The Groq row is the one to sit with. It is roughly ten times the throughput of a four-Mac cluster, and it caps context at 131,072 tokens — about an eighth of the advertised window. The million-token context and the fast inference are, for now, available in different places.
The Local-Cluster Alternative Is Already Cheaper
The thread supplied its own benchmark for what $11,000 of desk hardware should buy. colingauvin described running GLM 5.3 Flash on two DGX Sparks at about $7,000 each, getting 50–60 tokens per second at a full 1M context, and called it like having Claude at home. Same order of capital, roughly double to triple the throughput, a higher index score, and no million-token tradeoff. alexpotato said GLM 5.3 Flash is simply the daily driver. Meta's open release has to win against that, not against nothing.
The License Clause That Sets Your Price
Muse Spark Open ships under the Muse Model License, not MIT or Apache. Three terms matter in practice.
The $20 Million Trigger
Hacker News user seizethecheese read the license and surfaced the clause that drove most of the argument: a company above roughly $20 million in annual revenue that sells access to the model — model-as-a-service, in other words — has to negotiate a separate agreement with Meta. Below that threshold you can integrate it into a commercial product freely. This is not a restriction most readers will personally hit. It is a restriction every serious inference host hits, which is exactly why it shapes what you pay. Look at the hosted pricing again: Baseten, Fireworks AI, Together AI, Modal and Nebius Token Factory all charge the identical $1.25 input, $4.25 output, $0.15 cached input per million tokens. On a genuinely permissive open-weight model, that kind of price convergence is unusual; providers normally undercut each other. Here, every one of them is reselling under terms negotiated with the same counterparty. The license does not set a price, and the sources do not say Meta sets one — but it does mean nobody can route around Meta to compete on it.
Attribution and Use Restrictions
Products built on it must display "Built with Muse Spark Open" somewhere in the UI, documentation or settings. The model is not intended for anyone under 18. Prohibited uses include unlicensed professional practice and presenting outputs as human-authored. Techmeme's coverage notes derivative-model naming requirements and restricted-domain provisions comparable to Kimi's and GLM's licenses — so Meta is not an outlier here, which is itself the point. isityettime argued in the thread that terms like these disqualify the "open source" label entirely; this is open-weights, with conditions. The distinction is not pedantic when the conditions are what determine your bill.
What This Changes If You Are Deciding Today
If you wanted to run a frontier-adjacent model on your own hardware, this release does not help you, and the models that do help you are mostly Chinese. If you want a US-origin open-weight model you can audit, fine-tune and self-host inside your own compliance boundary, this is the strongest thing available and the deployment surface is broad enough to be credible. If you are just buying tokens, the honest comparison is against other hosted options on cost per completed task rather than per million tokens — bscphil's point in the thread was that real-world costs track context caching behavior more than headline rates, and the $0.15 cached-input figure is the number that will actually move your invoice.
Two things remain unresolved. The sources do not agree on what generation of the closed Muse Spark line this corresponds to — Meta's own post references 1.2, and the open question in the community is whether a Muse Spark 1.3-grade open release ever follows. And the license argument is live rather than settled: as long as every host charges the same rate, the clause that put them all at the same table will keep getting read aloud. Watch whether a sixth provider shows up with a different price. That will tell you more about how open this model really is than any benchmark will.
-EditorZ
Photo by Howard Bouchevereau on Unsplash
Post a Comment