On October 15, 2026, Meta Superintelligence Labs released Muse Spark Open. The official spec sheet is blunt: 816B total parameters, roughly 42B active per token, a sparse mixture-of-experts architecture, native input across text, images, video and audio, and a 1M-token context window. Meta calls it "the most capable US open-weight model to date," a company description rather than a third-party verdict.
The same launch material contains a sentence that is easy to scroll past: the model runs on a desktop, using two Mac Studios. Put 816B and "desktop" in one sentence and you have the real story. The weights were released. The barrier was not; it moved from a data center to your desk.
A note on scope: this site covered Muse Spark 1.1 on July 31, 2026, the closed model and Meta's first paid Model API. This article covers the open-weight release, its shared base model with Muse Spark 1.3, and what running it locally takes. The two pieces do not overlap.
1. The Spec Sheet
| Item | Value | Source |
|---|---|---|
| Total parameters | 816B | Official |
| Active per token | About 42B | Official |
| Architecture | Sparse MoE | Official |
| Context | 1M tokens | Official; a Hugging Face explainer cites 1,048,576, varying by runtime and configuration |
| Input | Text, images, video, audio | Official |
| Output | Text | Official |
| Base model | Shared with Muse Spark 1.3 | PyTorch official blog |
| Delivery | Open weights, not open-source code | Official |
First, the shared base model. Meta discloses no further training detail, but this is not a cut-down "light" variant built for a desk. The backbone is the 816B model.
Second, 1M context depends on runtime and configuration. The Hugging Face explainer cites 1,048,576 tokens and notes in the same piece that usable length varies with runtime and settings. Writing 1M is safe; treating it as a floor on any machine is not.
Third, audio is an input and text is the only output. This is not a speech model or a video generator; audio and video are things it understands, not things it produces.
Set that against Meta's open-weight history: Muse Glimmer 30B sat in the "one machine, have a go" range, 816B sits in the "plan the hardware" range. A 27-fold jump inside one product line is the heaviest engineering fact of this release.
2. 42B Active Is Not 42B Resident
This is the arithmetic that matters, and the reason an open-weight release can still be unrunnable for most people.
Sparse MoE routes each token to a small subset of the model's experts, so roughly 42B parameters take part in the matrix multiplications and compute per token lands near a 42B dense model. That is where the hope of speed comes from.
But "how much activates" and "how much must be stored" are different questions. The router can send the next token to any group of experts, so every expert must sit in memory waiting to be called; you cannot load only the popular ones or page them in like a browser. The full checkpoint must be resident, plus KV cache, activations and runtime overhead.
A rough order of magnitude, estimated from bytes and not an official figure: at BF16 with 2 bytes per parameter, 816B is about 1.6TB; at FP8 about 816GB; at 4 bits about 408GB. That is why the 1TB unified memory line matters. It is not headroom; it is the line between just enough and not enough. Above the quantized weights you still need room for context cache and runtime state.
Meta has not published the precision used for the desktop demo, the actual memory footprint, or the usable context length there. The numbers above explain scale only. They cannot be turned around to claim that two Macs run a full 1M-token context. That distance is exactly what most reposts leave out.
Fitting in memory is also not the same as being usable. Every MoE layer must read the selected expert weights out of memory before any math happens, so scattered routing eats effective bandwidth.
3. How Two Macs Become One
The official desktop path: two Mac Studios with M5 Ultra chips, 512GB of unified memory each, about 1TB combined, 1.2TB/s of bandwidth, and the latest Apple MLX. Meta's phrasing is that the two machines behave like one supercomputer with a terabyte of unified memory.
Bandwidth rather than compute is the point, because autoregressive decoding is usually bound by memory bandwidth: each generated token requires reading the slice of weights that token activates. How fast those weights can be read sets the ceiling on generation speed. Meta's blog says the same: 1.2TB/s is enough to read the small slice each token needs.
But "behaving like one machine" is not free. Inter-machine communication, weight sharding and round trips between layers all add latency. The official material gives the qualitative claim and no interconnect bandwidth, latency or tokens per second. Without those three numbers nobody can tell whether this does the work or merely gets it running. That is the biggest information gap in the launch. Two machines also mean two power supplies, two thermal envelopes and twice the failure points, and none of that shows up in a demo.
4. What Ran on the Desk
Two demos, both about finishing work rather than posting a score.
The first, driven by Muse Code, Meta's own agent tool, shipped a real iOS change to Ice Cubes, an open-source Mastodon client: a bug fix, a new feature, and a complete build, on-device.
The second drove DaVinci Resolve end to end, reviewing footage frame by frame, picking the strongest shots, cropping to vertical.
The choice is deliberate. The first maps to agentic coding: read code, change code, run a build, a long chain with little tolerance for error. The second maps to multimodal understanding plus tool use. Together they speak to whether long-horizon work closes out, not to a benchmark.
These are official demos, not independent reproductions. A demo proves the path exists, not that it exists on your machine. What it also cannot show is repeatability: one clean iOS change is a single sample, and agentic work is judged over dozens of attempts.
5. Ecosystem: Seven Engines, Four Quant Paths, One Helion
To judge whether an open-weight release was done seriously, read the runtime list before the benchmark table:
| Layer | Content | Source |
|---|---|---|
| Inference runtimes | SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp, Unsloth (7) | Official |
| Hardware validation | NVIDIA, AMD, Intel, Apple silicon | Official |
| Quantization | BF16 and NVFP4 (Meta), FP8 and MXFP4 (AMD), INT4 and MXFP4 optimized for Intel Crescent Island, GGUF (Unsloth) | Each vendor |
| Speculative decoding drafts | DFlash V2 in SGLang, DSpark in vLLM | Official |
| Post-training | TorchTitan, SGLang with Miles, NVIDIA NeMo AutoModel, Hugging Face TRL, plus hosted post-training APIs from Fireworks AI and Together AI | Official |
The four quantization paths decide whether the model lands on different hardware at all. BF16 and NVFP4 target NVIDIA's current generation, FP8 and MXFP4 target AMD, INT4 and MXFP4 are tuned for Intel Crescent Island, GGUF points at llama.cpp and consumer devices. Parameter count sets the ceiling; quantization and runtimes set how many people reach it. Note what the table implies: nobody had to wait for Meta to ship a container. The runtimes were ready the same day, which is what separates a weight drop from a platform.
Helion is a PyTorch-native high-level DSL that supplies auto-tunable GPU kernels inside SGLang. Vendor-reported gains: 1.03x to 3.95x at kernel level across NVIDIA and AMD, and 3% end to end on AMD MI350. Both are vendor measurements, not independent reruns.
Read those two together: up to 3.95x at kernel level, 3% end to end. Kernel gains reach the end only when kernels were the bottleneck, so when a launch claims a speedup, ask which layer the multiple belongs to.
The post-training row matters as much. TorchTitan, SGLang with Miles, NeMo AutoModel and TRL span pretraining frameworks through RL post-training, while hosted APIs from Fireworks AI and Together AI turn "modify the model" into "rent compute by the hour." Most teams will deploy a post-trained, narrowed version of these weights rather than the raw checkpoint, and whether that toolchain exists on day one says more than a leaderboard rank.
6. It Says It Is the Strongest; Its Own Chart Shows Losses
According to AI Daily Digest for October 17, 2026 on dev.to, relaying Meta's own comparison chart (second-hand, not an independent rerun), Muse Spark Open takes first place in six comparisons, including the tool-use benchmarks JobBench and MCP-Atlas.
The same chart shows where it loses: on terminal coding and long-horizon software engineering, measured by GDPVal-AA, it trails Kimi K3 and GLM 5.3.
That sits awkwardly against the positioning. Coding, agentic workflows and long-horizon tasks are exactly what the official material names as the target, and GDPVal-AA measures terminal coding and long-horizon software engineering. On its home turf, "the strongest" was caught by two open-weight models.
The practical reading: do not ask whether it is the strongest, ask which row your work falls into. Editing a repository of several hundred thousand lines over several days is a different capability from running a toolchain, and no ranking table separates them for you.
7. Open Weights Is Not Open Source: License, Distribution, Safety
The license is the part most often waved through. Muse Spark Open ships under the Muse Model License with an accompanying usage policy. Meta's own description is "limited commercial terms, which allow many developers to use the model for free."
That sentence should not be expanded by anyone but the license itself. What is permitted and what is restricted is a question for the LICENSE text; this article does not restate or interpret clause details, and does not judge whether a particular use falls inside the allowed range. Open weights means the weights can be downloaded, not that the license is OSI-style open source. By 2026 those two things have been conflated for far too long.
Distribution is broad: the meta-models/muse-spark-open collection on Hugging Face, plus Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted inference comes from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Anyone who does not want to build a machine has an API on day one. For a team weighing this release, that list is the practical shortcut: test through an API before buying hardware.
On safety, Meta says this release adds safety post-training compared with the August Muse Spark, making it more resistant to misuse when run standalone without system guardrails. Evaluation followed the Meta Superintelligence Scaling Framework and cleared the release threshold. Details are in the model card; the full evaluation report is "coming soon," meaning it is not out yet.
Closing: The Barrier Moved to the Desk, It Did Not Disappear
It genuinely lowers the barrier. Running an 816B-class model used to require a GPU room, cluster scheduling and an operations stack. Meta's official path is two maxed-out Mac Studios plus MLX, plug in and start.
But the barrier changed shape. The first wall is money: Meta has published no pricing tied to this demo, and this article will not guess; that it takes two machines already says a laptop will not do. The second wall is precision and configuration: with no published precision, memory footprint or usable context for the desktop demo, a reader cannot tell what they would get. The third wall is speed: with no tokens per second, there is no answer to "can I work on this."
A more accurate framing: Meta moved the open-weight barrier from "do you have a data center" to "will you pay for a desk." For a team that needs local deployment and cannot let data leave the intranet, that math may work. For an individual developer, the seven hosted providers are the more realistic route.
As for "made in America": according to a report by Global AI News, Meta's chief AI officer Alexandr Wang has said that most of the world's AI will run on open models and that those models should be made in America. That is second-hand, offered as a footnote; open weights has never been purely technical.
The line worth remembering: the 42B active figure buys speed. The 816B checkpoint is the memory bill. Meta put both in the blog; one got amplified, the other was folded into a sentence about two Mac Studios.
Related reading
- Meta Muse Spark 1.1: Zuckerberg Returns to X With a 1M-Context Agentic Model and Meta's First Paid API: The July predecessor and Meta's first paid API
- One Million Output Tokens: Which Model Finishes the Job: Who can actually hold a million-token context
- Local Coding Models: From 8GB GPU to 8-GPU Node: Where local models actually fit, from 8GB to eight cards