Meta Muse Spark Open: The 816B Open-Weight Model Anyone Can Run

Originally published on agent5.news.

For years, the biggest and most capable AI models have lived behind API paywalls, accessible only as a cloud service on someone else's hardware. The release of Meta's Muse Spark Open on October 15, 2026 pushes hard against that norm. With 816 billion total parameters, a one-million-token context window, and weights anyone can download, Meta and Hugging Face describe it as the largest open-weight model ever released by a US lab, a superlative from the releasing parties that has not been independently verified, and it is explicitly designed to run on hardware ranging from a pair of consumer-grade desktop computers to a rack of data-center GPUs. That combination of raw scale and genuine openness is why the AI community sat up and took notice.

What Muse Spark Open Actually Is

Muse Spark Open is the open-weight release of Muse Spark, a model that Meta first launched in April 2026 to power its own consumer AI products. The open version, published by Meta Superintelligence Labs, is a natively multimodal mixture-of-experts model with 816 billion total parameters, approximately 42 billion active per token, and a context window of up to one million tokens. Meta describes it as built specifically for coding and agentic work, excelling at long-horizon tasks, complex software engineering, and multi-step reasoning.

According to launch documentation, Muse Spark Open shares the same base model as Muse Spark 1.3, but received additional safety post-training to make it more robust as a standalone model without the system guardrails that surround the hosted product. In short: it is not a cut-down research toy, but a close relative of the production model Meta deploys to its own users.

The Architecture: Mixture of Experts, Explained Simply

The key to understanding Muse Spark Open is its mixture-of-experts (MoE) architecture. Think of MoE as a large committee of specialists rather than one generalist doing all the work. The model contains 256 experts inside each MoE layer, but only 8 of those experts are activated for any given token, selected by a routing mechanism called top-k routing. That means roughly 42 billion parameters do the actual computation on each token, even though the full model holds 816 billion.

The practical consequence is that per-token compute costs stay in the same range as a mid-sized dense model, even at frontier scale. What does not shrink is the memory requirement: every expert has to sit in GPU or unified memory and be ready to be called, because the router picks a different set for each token. The 816 billion parameters must be loaded in full, which is the central hardware challenge the model poses.

The full architecture, as documented in the model card, consists of 62 layers and a hidden dimension of 8,192. The vocabulary covers 202,048 tokens. On top of the language backbone sits a separate 1.8-billion-parameter Perception Encoder, which handles images, video, and audio; the model card lists 50 layers and a 4,096-dimensional output into the main model, though those specific sub-figures should be verified directly against the model card. That is what makes the model natively multimodal: it processes text, images, video, and audio as inputs and returns text.

Reasoning Effort: A Dial for How Hard the Model Thinks

One of the more distinctive engineering choices in Muse Spark Open is a parameter called reasoning_effort. Developers can set it to low, medium, high, or xhigh, and the model adjusts how much it reasons before responding. The default is xhigh, which Meta recommends for complex problem-solving, coding, and agentic tasks. Lower settings produce faster, shorter responses at reduced compute cost.

Importantly, reasoning cannot be fully switched off. The effort controls are not fixed output-token budgets but throttles on a process that is always running. That design choice reflects the model's primary purpose: it is built from the ground up as an agentic model that maps out steps, acts, observes, and self-corrects over long horizons, not just a model that produces single-shot answers.

The model also ships with draft model support for speculative decoding. The draft model proposes several tokens at once, and Muse Spark Open verifies them in a single pass, so each verification step yields multiple tokens instead of one. This is a standard technique for improving throughput without changing the model's outputs.

What the Benchmarks Show

Meta's model card compares Muse Spark Open, running at xhigh reasoning effort, against a set of other leading open-weight models including Inkling, Nemotron 3 Ultra, Beam, Mistral Large 4, Qwen3.8-2.4T-A95B, GLM 5.3, and Kimi K3.

On coding benchmarks, Meta reports a score of 60 on DeepSWE 1.1 and 58 on SWE-Atlas-QnA, two evaluations designed to measure real software engineering capability. On Terminal-Bench 2.1, the model scores 76.4. For context, the model card lists Kimi K3 scoring 85.0 and GLM 5.3 scoring 83.9 on that same Terminal-Bench test, meaning Muse Spark Open is competitive within the open-weight field but not at the very top of every leaderboard. On GPQA Diamond, a graduate-level reasoning benchmark, the model card reports 89.0; the same table shows Kimi K3 scoring 93.5 and GLM 5.3 scoring 91.7 on that benchmark, so Muse Spark Open trails on this measure as well. On Toolathlon Verified, which measures tool use, it reports 72.8.

A few caveats are worth keeping in mind. Benchmark numbers in this space are self-reported by labs and run under different harnesses and scaffolds, which makes direct comparisons tricky. Some rival scores in the model card comparison are also self-reported by those other labs. Readers should treat the numbers as directional signals, not as definitive rankings.

The Hardware Reality: What It Actually Takes to Run This

The words "anyone can run" require honest qualification. The weights are free to download. Running them is a different matter.

Meta's model card lists at least 700 GB of GPU memory for the NVFP4 quantized checkpoint, and 2 TB for the full BF16 precision version. The NVFP4 checkpoint, which quantizes only the experts, is reported by multiple sources as roughly 530 GB on disk, though the exact figure should be verified against the model card; the gap between checkpoint size and the minimum memory floor reflects activations, the perception encoder, and especially the KV cache needed for long-context work.

On the GPU server side, the NVFP4 build fits on four Blackwell Ultra GPUs or eight H200s. On the consumer side, Meta has specifically engineered the model to run on two Apple Mac Studio computers with M5 Ultra chips, each providing up to 512 GB of unified memory, linked over Thunderbolt 5 with RDMA. The pair together holds the full model in memory, though the 700 GB minimum memory floor accounts for activations and KV cache, meaning performance for long-context tasks on that configuration may be constrained. Meta tested a range of agentic workflows on that desktop setup, using a coding tool called Muse Code powered by Muse Spark Open, including shipping real iOS changes to Ice Cubes, an open-source Mastodon client, and driving DaVinci Resolve, the video editing software, end to end.

Weights are downloadable from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models.

Inference engine support at launch covers SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with optimizations for AMD, Apple, Intel, and NVIDIA silicon. Third-party reviews, including Eesel and Yotta Labs, report that Unsloth's 2-bit quantized build brings the memory requirement down to roughly 300 GB of combined memory, at some quality cost, though that figure should be verified against Unsloth's Hugging Face page.

The License: Freer Than Previous Meta Models, But Not Fully Open

Muse Spark Open ships under the Muse Model License, which is distinct from the Llama community license that governed earlier Meta releases. It is not Apache 2.0 or MIT.

The license requires a visible attribution notice in every product built on the model, specifically a "Built with Muse Spark Open" notice. For most developers, researchers, and businesses, that is the only meaningful restriction. There is one commercial trigger: model-as-a-service providers whose aggregate gross revenue exceeds 20 million dollars in any twelve-month period, according to the model card, must enter a separate agreement with Meta. Internal use, research, and products that embed the model are exempt from that revenue clause.

In practical terms, a startup building a coding assistant on top of Muse Spark Open, a researcher fine-tuning it for a domain-specific task, or an enterprise deploying it on private infrastructure all fall outside the MaaS revenue clause and can use the model freely with attribution.

Where Muse Spark Open Fits in the Bigger Picture

Muse Spark Open arrives at a moment when the gap between closed frontier models and the best open-weight alternatives has been narrowing steadily. The Muse series itself represents a significant inflection for Meta. Muse Spark was introduced in April 2026 as a major upgrade over the Llama 4 models; multiple sources note that Llama 4 drew criticism and that Muse Spark represented a strategic reset, though the characterization of that reception and the timeline of the Meta Superintelligence Labs division's formation have not been independently verified. The Muse name and the reorganization behind it are widely described as an effort to compete more directly at the frontier.

Muse Glimmer, a smaller model in the Muse family released earlier, preceded Muse Spark Open in the Muse lineup. Muse Spark Open is the full-size version, and it arrives in a week described as one of the most active single weeks for open-weight frontier models on record.

The broader significance is about control and economics. When a model of this capability class is closed, every inference dollar flows to the hosting lab. When the weights are open, teams can self-host, fine-tune for proprietary data, run air-gapped, and choose their own cost structure. For enterprises with compliance requirements or workloads large enough to justify the hardware investment, the economics shift meaningfully.

What to Watch Next

Several concrete developments are worth tracking in the weeks and months ahead. The Apple M5 Ultra Mac Studio with 512 GB of unified memory, which Meta specifically targeted for the two-machine desktop setup, was listed as coming in late October 2026, so real-world throughput numbers from that hardware configuration should emerge soon. Fine-tuning guides and reinforcement learning recipes are part of the official launch documentation, and community fine-tunes for specialized domains will arrive on Hugging Face as developers get access to the weights. The Muse Model License's MaaS revenue clause will face its first real interpretation tests as hosted inference providers finalize their commercial terms. And the benchmark conversation will continue: third-party evaluations run under consistent harnesses, rather than lab-reported numbers, will give a clearer picture of where Muse Spark Open actually sits relative to other frontier models.