JOURNAL / 2026.10.16
Meta releases Muse Spark Open: downloadable weights, huge memory needs and a custom license
The multimodal model opens another route to adapting and hosting agents. Its memory requirements and commercial terms determine who can actually benefit from that independence.
Meta released Muse Spark Open on October 15, with downloadable weights and announced availability through inference providers. The release lets developers retain a version of the model, modify it and choose where to run it. For teams that need control over that cycle, this is a material change even without accepting any of the announcement’s claims to leadership.
Its place within the family matters. The developer page describes it as a derivative of Muse Spark, released in August, with additional safety post-training. This is not the release of weights for Wonder, the model we covered this week. It changes Spark’s distribution and adaptability; it does not automatically make Meta’s entire latest offering downloadable.
The model card documents text, image, audio and video inputs, text output and a million-token context. That combination is relevant to an agent receiving a software bug alongside a screenshot and a recording: it can bring those pieces of evidence together in one model. A system still has to execute its tools.
Active size is not the memory requirement
Its 816 billion total parameters and roughly 42 billion active per token describe a mixture of experts. That selection reduces computation per token, but it does not let developers budget for a machine as though only 42 billion parameters needed housing. The other experts remain part of the model.
Meta documents configurations requiring at least 700 GB of GPU memory for NVFP4 and 2 TB for BF16. It also presents execution with MLX on two M5 Ultra Mac Studios, each with up to 512 GB of unified memory. That is a route to the desktop, but at a scale far removed from an ordinary laptop. The commercial description of “local” covers very different machines.
There are compression alternatives. The llama.cpp team’s repository lists 470 GB for Q4_K_S and 304 GB for Q2_K, plus separate perception components. Those are file sizes, not our measurements of total memory. The inference engine, context cache and concurrency require additional headroom; nor should we transfer scores from a different precision to these versions without checking.
That distinction changes the building decision. A team can maintain its own version on controlled servers without every update depending on an API provider. However, savings per request must cover memory, operations and actual utilization. For a lightly used application, hosting this model could buy independence at a cost that is hard to justify. For continuous workloads or specialized adaptations, the calculation deserves investigation.
What openness permits and what remains to be demonstrated
The Muse Model License permits modification, fine-tuning and distillation, with attribution and usage-policy obligations. It requires notifying Meta and signing a separate agreement to offer the model or derivatives as a service when the licensee and its affiliates’ aggregate gross revenue exceeds $20 million over twelve consecutive months. The clause excludes internal use and the licensee’s own products that are not such a model service. “Open weights” describes meaningful availability here, with conditions worth reading before designing a business.
I would not treat Meta’s ranking as proof of universal superiority either. Its methodology report combines internal runs, evaluator results and figures reported by other providers. On DeepSWE, which checks extended software-engineering tasks using verifiers and regression tests, models use different harnesses. The comparison provides guidance about evaluated systems; it does not isolate each model’s contribution within an identical integration.
My view is that the more lasting contribution may be the ability to maintain comparable experiments as commercial offerings change. Retaining weights, configuration and adaptations makes it possible to investigate a regression much more precisely. That possibility also requires recording quantization and tools: keeping only the model’s name would leave out part of its behavior.
Before crediting it with a practical advantage, I would ask for repeatable tests of complete tasks in a configuration the team can sustain, including errors, reviews and cost. The release opens that investigation to third parties. Its value will grow if it produces adaptations and inspectable results outside Meta, alongside demonstrations on exceptional hardware.
Primary sources consulted on October 16, 2026. We have not run the model or experimentally verified its requirements or performance.