Meta published open weights for Muse Spark Open on October 15, a mixture-of-experts model with 816 billion total parameters and 42 billion active per forward pass. The announcement, posted to research.meta.ai, lists a 1 million token context window, multimodal input, and a focus on coding and agent workloads. Meta says the model runs locally on two Apple M5 Ultra chips via MLX, and that both weights and APIs are available now.
That local-inference detail is the part worth sitting with. A model this size running on consumer-adjacent hardware, rather than a rack of data center GPUs, changes who can host it and where. Teams that previously ruled out self-hosting large models on cost or infrastructure grounds now have a documented reference configuration to evaluate against their own latency and throughput needs.
Meta's announcement confirms the parameter counts, the context length, the multimodal capability, and the two-chip MLX deployment path. It does not state throughput numbers, power draw, latency under load, or pricing for the API option. It also does not disclose fine-tuning support, rate limits, or how the 42B active parameters behave under sustained concurrent requests. Those are the figures an operations team would need before treating this as a production dependency rather than a lab experiment.
For platform operators, the open-weights release itself is the operational event, separate from any API Meta also offers. Self-hosted models remove a vendor status page from the dependency chain but add a new one: your own hardware. Anyone planning to run Muse Spark Open on M5 Ultra hardware needs to check memory headroom against the 816B total parameter footprint, even with only 42B active at inference time, because MoE architectures still require the full weight set resident somewhere in the serving path. Expect routing logic, not raw compute, to be the first thing that breaks under real traffic.
On-call engineers evaluating the hosted API path face a different checklist. Confirm whether existing alerting and rate-limit assumptions, built for smaller context windows, still hold at 1 million tokens, since a single long-context request can now consume resources equivalent to many short ones. Check timeout configurations against expected latency for multimodal and agent-style calls, which tend to run longer than simple text completion. None of this is disclosed by Meta's announcement, so it has to be measured locally before the model sits anywhere near a production incident path.
The practical split is between teams that self-host and teams that call an API, and Meta's announcement supports both without specifying operational limits for either. Two M5 Ultra chips is a concrete, named configuration, not a minimum or a guarantee. Capacity planning starts there.