Meta releases 816B-parameter Muse Spark Open

Meta Superintelligence Labs has released Muse Spark Open, a natively multimodal mixture-of-experts model with 816 billion total parameters, 42 billion active parameters per token, and a context window of up to 1 million tokens. Meta describes it as the most capable open-weight model yet released by a US lab, with an emphasis on coding, software engineering, tool use, and long-running agent workflows.

The weights are available through Hugging Face, while hosted providers and major inference frameworks supported the model at launch. Independent evaluations have yet to verify Meta’s capability claims.

816B parameters, 42B at work

A sparse mixture-of-experts model routes each token through a selected group of parameter blocks, called experts. Muse Spark Open activates about 42 billion parameters for each token while retaining the broader capacity of an 816 billion-parameter model. This lowers computation, although a self-hosted deployment still needs enough storage and memory for the full checkpoint.

The 1 million-token limit could let an application supply a large repository, a lengthy tool-call history, or extended multimodal input in one request. Actual performance will depend on the serving engine, available memory, prompt composition, and the model’s ability to retrieve information across long contexts. Large prompts also increase prefill time and key-value cache requirements.

Two Macs need careful math

Meta says Muse Spark Open is the first model in its class to run on two Apple M5 Ultra Mac Studios using MLX. Each system can provide up to 512 GB of unified memory, giving the pair a theoretical aggregate of 1 TB. MLX must shard the weights and coordinate inference across the two machines because their memory pools remain separate.

Sparse activation reduces the parameter data used for each token, but memory bandwidth, interconnect performance, cache size, and routing overhead still affect throughput. Meta cites up to 1.2 TB/s of memory bandwidth per M5 Ultra. Its announcement does not provide enough end-to-end latency or tokens-per-second data to compare the setup with GPU servers.

Meta reports running two on-device demonstrations with Muse Code. One modified and built Ice Cubes, an open-source Mastodon client for iOS. The other controlled DaVinci Resolve, reviewed footage frame by frame, selected clips, and produced vertical crops. These examples show the intended workflow, though they are vendor demonstrations rather than standardized agent benchmarks.

Serving support arrives with the weights

Meta also ships auto-tunable Helion GPU kernels for SGLang. Its tests report kernel-level speedups ranging from 1.03× to 3.95× on NVIDIA and AMD GPUs, plus a 3% end-to-end improvement on an AMD MI350 system. The smaller application-level gain indicates that kernel execution accounts for only part of total serving time. Results will vary with hardware, batch size, context length, and quantization.

Read the custom license before deployment

Hosted pricing is set by each inference provider, so Muse Spark Open has no single API rate. Meta distributes the weights under the custom Muse Model License, which includes limited commercial terms and allows many developers to use the model without a license fee. Teams planning production use should review the clauses covering commercial scale, redistribution, modifications, acceptable use, and hosted access.

Where the model should fit

Meta has tuned Muse Spark Open around workloads that benefit from large working memory, multimodal input, and repeated tool calls:

• Long-running agent trajectories that must preserve information from early steps
• Software engineering tasks spanning multiple files or repositories
• Multi-step reasoning with tools, code execution, and external services
• Applications combining code with images, video, or audio

Independent testing remains the gap

Meta’s release includes benchmark charts and a model card with safety results, but a full independent evaluation report is not yet available. Meta says a more complete report will follow. Until outside researchers publish results on SWE-bench, Terminal-Bench, long-context retrieval, multimodal tasks, and agent reliability, the company’s performance claims remain provisional.

Production evaluations should measure more than benchmark accuracy. Relevant checks include time to first token, output throughput, memory use at long context lengths, quantization loss, tool-call accuracy, prompt-injection resistance, and completion rates for full agent trajectories.

What changes for agent builders

1. A larger US open-weight option: The release adds an 816B-parameter model from a US lab to a field where Chinese labs have supplied many of the largest recent open-weight systems.
2. A wider local deployment envelope: Two high-memory workstations can host the model without a data-center rack, subject to practical limits on cost, speed, cooling, and interconnect performance.
3. Broad launch support: API providers, cloud catalogs, inference engines, and fine-tuning tools supported the model when the weights became available.

Teams building coding agents can now compare a single long-context model with architectures that route tasks among several smaller models. The deciding measurements will be task completion rate, latency, infrastructure cost, and reliability across long tool-driven workflows.