Meta's Muse Spark Open: What an 816B Download Actually Gets You
Synergised Consulting Ltd
Last Updated: 17 October 2026
Meta released the weights of Muse Spark Open on 15 October 2026 and called it the most capable US open-weight model to date. The claim is true in the narrow sense and unhelpful in the one that matters: on Artificial Analysis's independent index it scores 39, behind the Chinese open-weight models many businesses already run. What the release changes is who controls the weights, not which model tops the table.
What Muse Spark Open actually is
Muse Spark Open is an open-weight mixture-of-experts language model, released for download and served by hosted providers from day one, built for coding and agentic work. According to Meta's release post [1], it has 816 billion total parameters with 42 billion active per token, a context window of 1 million tokens, and native handling of text, image, video and audio input. The Hugging Face welcome post adds the shape behind those numbers [3]: 256 experts per layer with 8 selected per token, which is what keeps the compute cost per token near the smaller figure. Weights are available in BF16 and NVFP4 formats [4], under Meta's Muse Model License rather than an open-source licence [2], and the model card recommends deploying it behind additional guardrails rather than standalone [2]. It is served from launch by Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI [1].
The detail that matters for agents
According to vLLM's day-zero support post [4], Muse Glimmer 30B, the small model suited to running locally, was distilled from Muse Spark Open, and the small model inherits its text-handling layer from the larger one. For a business running a small local model today, that matters more than the headline size: the large model and the small one are built as a pair, so work proven against one transfers to the other with less friction. The vocabulary is 202,048 tokens [2] and the embedding layers account for roughly 1.66 billion parameters each [3], details that matter to whoever serves the model and not to whoever chooses it.
What the independent numbers say
On independent measurement, Muse Spark Open is a strong model that is not the leader, and the gap is widest exactly where agent work happens. According to Artificial Analysis [5], it scores 39 on the Intelligence Index, well above the median of 19 for comparable open-weight models, and the evaluation runs found it notably fast and relatively concise, generating 110 million output tokens against a median of 130 million [5]. But the same index places GLM-5.3 at 45 [6] and Kimi K3 at 44 [8], both open-weight and both already servable, and closed models sit higher still: Claude Opus 5.5 leads the index at 58 [7].
Measure | Muse Spark Open | GLM-5.3 (Max) |
|---|---|---|
AA Intelligence Index | 39 | 45 |
Terminal-Bench 4.0 | 8% | 42% |
AutomationBench-AA | 50% | 62% |
Humanity's Last Exam | 44% | 42% |
SciCode | 58% | 59% |
According to Artificial Analysis's comparison data [6], the pattern in the table is consistent: the knowledge-heavy evaluations are close, and the agentic ones are not. Terminal-Bench 4.0, which measures multi-step terminal work of the kind coding agents do, shows the widest gap on the board.
Reading a "most capable" claim
The qualified phrase does real work here. "Most capable US open-weight model" is accurate against a geography filter, and Meta's own launch table shows the same 39 against 45 for GLM-5.3 [6]. A founder reading only the headline would reasonably conclude the best downloadable model had changed hands. It had not. The honest summary is that Meta moved from having no frontier-scale open weights to having a credible one, which is a real shift in supply, stated without claiming a shift in rank.
What it is not for
Muse Spark Open is not a model a small business should plan to self-host, and on today's evidence it is not the one to pick as an agent's reasoning engine even when rented. The first limit is physical: 816 billion parameters in BF16 format implies a multi-GPU server commitment that no founder-operated business needs to make, and every practical route to using it, from Meta's own launch partners onward [1], is hosted serving of someone else's hardware. The second limit is measured: at 8% on Terminal-Bench 4.0 against 42% for GLM-5.3 [6], it is currently the weakest of the leading open-weight options at exactly the benchmark that simulates agent work, so choosing it for an agent stack on the strength of the release would mean trading measured capability for weight provenance. The third is licensing: the Muse Model License is a click-through with conditions [2], not the permissive licence some "open weight" coverage implies, so anyone embedding it in a product should read the terms rather than assume Apache-style freedom.
The worked choice
Picture the actual decision. An owner has been running agents on GLM-5.3 through a hosted endpoint, and a release headline says a bigger, newer model is now downloadable. The temptation is to read "816 billion parameters" as an upgrade signal. The arithmetic says otherwise: the incumbent scores higher on the independent index [6], the newcomer cannot be self-hosted at this size, and nothing in the release changes the incumbent's price or terms. The correct action is to note the release, record where the new model sits on the index, and revisit when independent agentic scores exist for it. It costs minutes and saves a migration nobody needs.
What this release actually changes
The change is in who controls usable weights, and that has value even when the model is not the leader. According to Meta's release post [1], the weights went to Hugging Face, Amazon SageMaker, Google Cloud Model Garden and Microsoft Foundry on day one, with seven hosted providers serving immediately. A credible, US-originated, frontier-scale option now exists in the download tier, and whichever model a business actually runs, its vendor now competes against a downloadable alternative that did not exist this time last month. Pricing and terms for everything in this tier just got a reference point they lacked.
In the discovery conversations we run, model questions rarely arrive as model questions; they arrive as a headline the owner has read, and the work is turning it back into a measured choice. The useful habit is the one this post applied to Muse Spark Open: check the independent index before the press release, check what the licence actually permits before the download, and write down what you found. That note is the difference between choosing a model and being marketed to.
Sources
- [1] Meta AI Research, "Introducing Muse Spark Open" (15 October 2026):
- [2] Meta, Muse Spark Open model card, Hugging Face (15 October 2026):
- [3] Hugging Face, "Welcome Muse Spark Open" (15 October 2026):
- [4] vLLM, "Day 0 support for Muse Spark Open on vLLM" (15 October 2026):
- [5] Artificial Analysis, "Muse Spark Open (Xhigh)" model page (data as of 16 October 2026):
- [6] Artificial Analysis, "Muse Spark Open (Xhigh) vs GLM-5.3 (Max)" comparison (data as of 16 October 2026):
- [7] Artificial Analysis, "Artificial Analysis Intelligence Index v4.3.2" (data as of 16 October 2026):
- [8] Artificial Analysis, "Muse Spark Open (Xhigh) vs Kimi K3 (Max)" comparison (data as of 16 October 2026):
