# Meta releases Muse Spark Open weights and serving guides > Meta released Muse Spark Open weights on October 15. Its BF16 and NVFP4 checkpoints, hosted APIs and Muse Model License give developers several routes with distinct limits. By BIG CHANGE Editorial Published: 2026-10-16T06:43:13.531Z Updated: 2026-10-16T06:43:13.531Z Canonical: https://bigchange.ai/blog/meta-muse-spark-open-weights-serving-license ![Conceptual charcoal illustration of an open server cabinet on the left and a separate wall-mounted network connection on the right.](https://bigchange.ai/api/media/file/muse-spark-open-serving-choice-hero-v1.png) AI-generated conceptual illustration by BIG CHANGE; no Meta equipment, actual deployment or BIG CHANGE test is depicted. Meta released the weights for Muse Spark Open on October 15, 2026. A team can now [download Meta's BF16 or NVFP4 checkpoints](https://dev.meta.ai/docs/muse-spark-open/get-the-model), serve a supported format on its own hardware, or call the model through a launch partner. The choice matters: the full checkpoint is about 1.65 TB of weights, the NVFP4 checkpoint about 537 GB, and a hosted endpoint has its own access and price terms. BIG CHANGE checked the launch documents, model card, serving guide and license. We did not download or run the model. Meta describes Muse Spark Open as a sparse mixture-of-experts model with 816 billion parameters in total and about 42 billion active per token. It accepts text, images, video and audio and produces text. The [model documentation](https://dev.meta.ai/docs/muse-spark-open) specifies a 1,048,576-token maximum context. That is a model limit, not a promise that every provider or local configuration will allocate that many tokens. ## The big change - **What changed:** Meta published downloadable Muse Spark Open weights under its Muse Model License, alongside documented routes for self-hosting, partner APIs and post-training. - **Why it matters:** Developers can choose where inference runs and which checkpoint or provider fits their workload. The released recipes make that choice more concrete, while memory, context, price and access still depend on the route. - **What to watch:** Independent tests of the released checkpoint, provider-specific context and pricing, and the behavior of the smaller quantizations on real tasks. Meta's demonstrations and evaluations do not establish those results for another deployment. ## Weights, hardware and context Meta's [download guide](https://dev.meta.ai/docs/muse-spark-open/get-the-model) names two first-party repositories on Hugging Face: `meta-models/Muse-Spark-Open-816B-A42B` for BF16 and `meta-models/Muse-Spark-Open-816B-A42B-NVFP4` for the smaller checkpoint. The BF16 weight files are about 1.65 TB; NVFP4 is about 537 GB. Those are storage sizes, before key-value cache, activations and runtime overhead. Meta's serving examples put BF16 on eight B300 GPUs and NVFP4 on four B300 GPUs. Its [hardware table](https://dev.meta.ai/docs/muse-spark-open/hardware-platforms) also describes AMD Instinct configurations with vLLM, an Intel Xeon 6 INT4 route with SGLang, and an Apple MLX configuration using two M5 Ultra Mac Studios with 512 GB of unified memory each, or four with 256 GB each. These are named recipes, not a general compatibility guarantee. For a workstation with much less memory, the [deployment guide](https://dev.meta.ai/docs/muse-spark-open/deploy) points to third-party GGUF quantizations. Its Unsloth planning range is roughly 300–315 GB of combined RAM and VRAM for a dynamic 2-bit version. That is a different checkpoint and runtime from Meta's BF16 and NVFP4 downloads. The example Unsloth command starts with a 32,768-token context to control memory use. A team should size its actual context and parallel requests against available memory rather than read the model's one-million-token ceiling as a default serving setting. ## Choose where requests run The [serving guide](https://dev.meta.ai/docs/muse-spark-open/deploy) documents SGLang, vLLM and TensorRT-LLM server paths, a Transformers Python path, MLX for the specified Mac Studio arrangement, and llama.cpp and Unsloth with their GGUF variants. It lists OpenAI-compatible HTTP endpoints for several server runtimes; Transformers uses a Python interface. Each recipe names its checkpoint, parallelism and runtime settings. Selecting a different quantization or accelerator calls for that route's own recipe. Meta also lists launch-day hosted access through Baseten, Fireworks AI, Modal, Nebius, OpenRouter and Together AI. Groq's route is described as a limited preview, so access there depends on an eligible organization. The [partner section](https://dev.meta.ai/resources/blog/build-with-muse-spark-open#hosted-inference) distinguishes serverless and dedicated offerings. It does not establish one price or context allocation across all providers. A hosted API uses the provider's credentials and billing; local inference requires the checkpoint, substantial storage and compatible compute. Meta's download guide says an offline local run needs no model API key after the weights are obtained. For a first evaluation, developers can use a partner endpoint to test their own prompts and tool calls, or follow the exact local recipe for their checkpoint and hardware. Meta's [model card](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) warns that the model can produce inaccurate answers and errors in multi-step reasoning. It recommends system-level safeguards for agents with tools. Neither a launch demonstration nor the company's reported benchmarks show that an agent will reliably complete a particular team's workflow. ## The license has conditions The [Muse Model License](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B/blob/main/LICENSE.md), dated October 15, permits use, modification and distribution subject to its conditions and attached usage policy. It requires the license and policy to accompany copies and derivatives, and a prominent “Built with Muse Spark Open” notice for products or services using the software or derivatives. A model-as-a-service provider whose aggregate gross revenue with affiliates exceeds $20 million in any consecutive 12 months must notify Meta and enter a separate written agreement before offering the model or derivatives in that business. The license expressly excludes internal use and commercial products that use the model as a component without offering a model-as-a-service product from that particular requirement. Teams planning a deployment should read the current license and policy for their use. Downloadable weights do not make the release unrestricted open source, and hosted inference or hardware is not generally free. Meta released this model three days after [introducing the hosted Muse Wonder family](https://bigchange.ai/blog/meta-muse-wonder-agent-model-access). Spark Open gives teams a separate weight-download and deployment decision. The practical question now is which documented route can support the team's intended inputs, context, budget and review process. That remains a configuration and evaluation choice, not a result BIG CHANGE has measured. ## Sources & further reading - [Meta AI Research release, October 15, 2026](https://research.meta.ai/blog/muse-spark-open/) — release date, model description, partner list and company demonstration. Its capability and evaluation statements are Meta's. - [Meta developer launch guide, October 15, 2026](https://dev.meta.ai/resources/blog/build-with-muse-spark-open) — route-by-route runtime, checkpoint, hosted provider and post-training documentation. Its benchmark and early-user statements remain attributed to their sources. - [Meta's model, download, hardware and deployment documentation](https://dev.meta.ai/docs/muse-spark-open) — exact inputs and output, footprint, recipes, prerequisites and provider access; read each linked route's current settings before deployment. - [Muse Spark Open model card](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) — model limits, Meta's evaluations and stated limitations. BIG CHANGE did not run or independently reproduce them. - [Muse Model License and usage policy, October 15, 2026](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B/blob/main/LICENSE.md) — applicable redistribution, attribution and model-as-a-service terms. This summary is not a substitute for the full text. ## Sources - [Meta AI Research: Introducing Muse Spark Open](https://research.meta.ai/blog/muse-spark-open/) — October 15, 2026 release; model description, hosted partner listing and company demonstration. Meta claims are attributed. - [Meta AI Developers: Build with Muse Spark Open](https://dev.meta.ai/resources/blog/build-with-muse-spark-open) — October 15, 2026 guide to runtime, quantization, hosted provider and post-training routes. - [Meta Model API: Muse Spark Open](https://dev.meta.ai/docs/muse-spark-open) — Inputs and output, model context, checkpoint and serving overview. Live documentation checked October 16, 2026. - [Meta Model API: Get the model](https://dev.meta.ai/docs/muse-spark-open/get-the-model) — BF16 and NVFP4 checkpoint repositories, weight footprints and prerequisites. - [Meta Model API: Hardware platforms](https://dev.meta.ai/docs/muse-spark-open/hardware-platforms) — Named NVIDIA, AMD, Intel and Apple configurations. These are route-specific. - [Meta Model API: Run inference](https://dev.meta.ai/docs/muse-spark-open/deploy) — Runtime APIs, partner access, context choices and Unsloth memory planning. - [Muse Spark Open model card](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) — Meta's model specifications, safety evaluation and explicit limitations. BIG CHANGE did not reproduce them. - [Muse Model License and usage policy](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B/blob/main/LICENSE.md) — October 15, 2026 release; conditions for copying, product notice, and model-as-a-service revenue provision.