What changed
On October 15, 2026, Meta released Muse Spark Open, an 816-billion-total/42-billion-active-parameter mixture-of-experts model with 1-million-token inference context and text, image, video and audio inputs (text output). Unlike Meta's existing hosted Muse Spark 1.3 and newly announced Muse Wonder models, this release includes downloadable BF16 and NVFP4 weights. Meta lists hosted inference through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, and software support including vLLM, SGLang, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth. The Muse Model License permits commercial use and derivatives but requires attribution and a separate agreement before large model-as-a-service operators exceeding $20 million in aggregate affiliated gross revenue over any consecutive 12 months offer the model or derivatives. Together AI lists $1.25 per million input tokens, $4.25 output and $0.15 cached input.
Why it matters
This gives builders a new way to evaluate, fine-tune and serve a very large US-developed agent model without relying solely on a proprietary endpoint. The 42B-active figure describes per-token compute, not total memory: Meta's model card calls for at least 700GB of VRAM for NVFP4 and 2TB for BF16, so ordinary local machines are not a realistic hosting target. Meta demonstrates operation across two 512GB-memory M5 Ultra Mac Studios, which is a specialist setup. Hosted APIs lower the entry barrier, but their prices and terms differ by provider. Indie SaaS teams can generally embed the model in their own non-model-as-a-service products without triggering the $20M clause; resellers and inference businesses must pay particular attention to the separate-agreement requirement.
Downloadable weights, not merely an API preview
Meta released 816B-total/42B-active Muse Spark Open weights on October 15, 2026. The model has a 1M-token context window and accepts text, images, audio and video, with text output. It is built for long-running coding agents, tool calls and multimodal work. This is distinct from the September Muse Spark 1.3 hosted-model update and October 12 Muse Wonder access announcement.
The memory bill changes the self-hosting calculation
Meta's official model card says NVFP4 inference requires at least 700GB of GPU memory, versus about 2TB for BF16. It also describes an MLX deployment using two Apple M5 Ultra Mac Studios with up to 512GB unified memory each. An 816B-total mixture-of-experts model still has to store its weights even when only 42B parameters are active per token. Smaller operators are more likely to begin with a hosted provider than local hardware.
The licence permits commercial work, with explicit conditions
The Muse Model License permits use, modification, fine-tuning, distillation, redistribution and commercial applications. It requires preserving the licence and usage policy and prominently displaying 'Built with Muse Spark Open' in products, services or documentation. An operator whose own and affiliates' aggregate gross revenue exceeds $20M over any consecutive 12 months must notify Meta and obtain a separate agreement before offering the model or derivatives as a model-as-a-service API. The clause expressly excludes internal use and using the model to power a non-model-as-a-service product. This is not a universal $20M cap on all commercial uses.
Hosted inference offers a lower-capital entry path
Meta lists multiple launch-day inference partners. Together AI's October 15 price is $1.25 per million input tokens, $4.25 per million output tokens and $0.15 per million cached input tokens for its serverless service. These are Together-specific rates, not universal Meta list prices. Developers should verify each provider's context, latency, retention, multimodal coverage and service-level limits before comparing unit prices.
Benchmarks are informative but not independently settled
Meta's model card includes detailed benchmarks with its own generation settings, including xhigh reasoning effort. These vendor evaluations should not be treated as proof of a universal performance lead or compared casually against competitors tested under different harnesses and budgets. Independent reporting confirms the release and unusual Apple Silicon collaboration, not a replicated head-to-head coding win.
Operational and safety boundaries
Meta warns that the released checkpoint can still fail on complex multi-step tasks, underrepresented languages and low-resolution video, and recommends guardrails and human confirmation for irreversible agent actions. The model's additional safety post-training and reported refusal rates are Meta's own measurements. Teams should evaluate their actual tool-use workflows, license compliance and data-handling policies before production deployment.