Muse Spark Open explained: what Meta's 816B open-weight model costs to run, and what its license lets you do

Meta Superintelligence Labs released the open weights of Muse Spark Open, an 816 billion parameter mixture-of-experts model with 42 billion active parameters, on October 15, 2026. The Muse Model License is free for almost everyone, but a company with more than $20 million in total yearly revenue that sells access to the model through an API must sign a separate deal with Meta, and every product using it must display "Built with Muse Spark Open". The full BF16 weights are 1.64 TB, the 4-bit NVFP4 build is 533 GB on Hugging Face, and on OpenRouter the model costs $1.25 per million input tokens and $4.25 per million output tokens.

Muse Spark Open explained: what Meta's 816B open-weight model costs to run, and what its license lets you do Meta released the weights of Muse Spark Open, an 816 billion parameter model, on October 15. This page reads the license line by line, works out how much memory it really takes from the actual file sizes on Hugging Face, sets Meta's own benchmark table against the open models that beat it, and sorts who can use it free and who needs a deal with Meta. This explains reporting by Meta Superintelligence Labs, "Muse Spark Open" Meta AI research blog, October 15, 2026 https://research.meta.ai/blog/muse-spark-open . Read the original first: https://research.meta.ai/blog/muse-spark-open https://research.meta.ai/blog/muse-spark-open In one minute - Muse Spark Open is a mixture-of-experts model with 816 billion parameters, of which 42 billion work on each token. It reads text, images, video and audio, writes text, and has a 1 million token context window. - The license is free for almost everyone. The big catch is narrow: a company with more than $20 million in total yearly revenue that sells access to the model through an API must sign a separate deal with Meta first. Using it inside your own product does not count. - Every product that uses it, or any model trained on its outputs, must show "Built with Muse Spark Open" somewhere a user can find it, and must follow Meta's usage policy, which bans things like writing malware and running critical infrastructure. - The full BF16 weights are 1.64 TB. The 4-bit NVFP4 build is 533 GB on Hugging Face, which fits on two 512 GB M5 Ultra Mac Studios or four Blackwell Ultra GPUs. A 2-bit Unsloth build is 311 GB. - In Meta's own table it leads on agent tests like MCP-Atlas 89.2 and JobBench 61.8 , but trails GLM 5.3, Kimi K3 and Qwen3.8 on the Artificial Analysis Intelligence Index. On OpenRouter it costs $1.25 per million input tokens and $4.25 per million output tokens at every listed provider. What Meta released On October 15, 2026, Meta Superintelligence Labs released Muse Spark Open, which the blog calls the most capable US open-weight model to date. Open weight means you can download the trained model file and run it yourself, on your own machines. The model card gives the shape. It is a mixture-of-experts transformer with 816B total parameters and 42B active. It has 62 layers and 256 experts, and 8 experts work on each token. A separate 1.8B parameter perception encoder handles images and video. Input can be text, image, video or audio, and output is text. The context window is 1M tokens. Mixture of experts is why both numbers matter. Every one of the 816 billion parameters must sit in memory, so memory is set by the big number. But each token only runs through about 42 billion of them, so speed is closer to that of a 42B model. The card says the open version got extra safety training so it holds up "as a standalone model without system guardrails", and Meta still recommends deploying it inside a system with its own guardrails. The license, clause by clause The Muse Model License reads like the MIT license with three conditions bolted on. The grant is broad: free of charge, you may use, copy, modify, fine-tune, distill, merge, publish, sublicense and sell copies. Meta also says it claims no ownership of outputs. Condition 1, the usage policy. You must include the license and the usage policy with every copy, and your use must follow the policy, which is attached as Exhibit A. Condition 2, attribution. You must "prominently display" the words Built with Muse Spark Open in the user interface of each product that uses the model, or in its documentation, website or settings menu. A line in your settings or docs page is enough. Condition 3, the revenue line. If you give third parties access to AI models through an API or similar interface, which the license calls Model as a Service, and you and your affiliates together have more than $20 million in gross revenue in any 12 months, you must notify Meta and sign a separate agreement before you offer this model that way. Read the exceptions closely, because they are wide. Section 3 does not apply to internal use, to research, to any commercial use that is not Model as a Service, or to using the model to power features of your own product. So a large company running it on its own help desk, or inside its own app, needs no deal. A cloud host or an API reseller over $20 million does. Note what the $20 million measures. It is the total gross revenue of you and your affiliates, not the revenue the model earns. A company with $25 million in sales from something else entirely crosses the line the day it opens a Muse API to customers. The definition of Derivatives reaches further than most readers will expect. It means any AI model created, trained, modified, distilled or otherwise improved using the Software "or any outputs or results thereof". If you train a small model on Muse Spark Open's answers, that small model is a Derivative, and the attribution rule and the usage policy come with it. The usage policy items that will matter at work Most of Exhibit A bans what you would expect, such as violence, fraud and child abuse material. A few items are worth reading twice if you build tools for a business: - Not intended for anyone under 18. Deployers who might reach minors must assess and reduce the risk themselves, the model card says. - No creating malicious code, malware or viruses, or anything that could disable or impair a website or computer system. That is a gray zone for security teams who write exploit tests. - No use for operation of critical infrastructure, transportation technologies or heavy machinery. - No military, warfare, nuclear or espionage uses, or anything under ITAR export rules. - No unlicensed practice of a profession, including financial, legal and medical practice. - No presenting the model's output as human generated, and you must disclose any known dangers of your AI system to end users. Getting the weights is gated too. Hugging Face shows the repository with manual approval, and the form asks for your full legal name, date of birth, country, affiliation and job title, and records your IP location. How much memory it really takes Meta and Apple give three different numbers for the 4-bit build. Apple's MLX post says the NVFP4 version is about 493GB. Meta's developer guide says about 530 GB. The model card says you need a GPU setup with at least 700GB of VRAM for NVFP4 and 2TB for BF16. We added up the file sizes Hugging Face reports for each repository on October 16: - BF16, the full precision weights: 1,644 GB. - NVFP4, Meta's 4-bit build: 533 GB. - llama.cpp GGUF from ggml-org: Q8 0 872 GB, Q4 K S 470 GB, Q2 K 304 GB. - Unsloth Dynamic GGUF: UD-Q4 K XL 506 GB, UD-Q2 K XL 311 GB, UD-IQ1 S 176 GB. So the 533 GB figure is the file. The 700 GB figure is the file plus working room, and most of that working room is the KV cache, the memory a model keeps for every token of context it has read. The card gives enough detail to estimate the cache. It lists 62 layers, 16 key-value heads and a head size of 64. One token then stores 2 x 62 x 16 x 64 = 126,976 numbers. At 2 bytes each, that is about 254 KB per token. A 128,000 token conversation needs about 33 GB of cache, and the full 1 million tokens needs about 254 GB. That is our arithmetic, and it assumes every layer keeps a full cache in 16-bit numbers. Put the two together and Meta's 700 GB makes sense: 533 GB of weights plus about 167 GB, which holds roughly 650,000 tokens of 16-bit cache. If you want the full 1M window at 16 bits on one box, plan for about 790 GB. What fits where, using Meta's own supported setups: - Four Blackwell Ultra GPUs or eight H200s, with SGLang, vLLM or TensorRT-LLM, for NVFP4. - Two M5 Ultra Mac Studios with 512GB each, or four with 256GB each, linked with Thunderbolt 5 and running MLX. That is about 1 TB of unified memory, so NVFP4 fits with room for cache. - About 300 GB of combined memory for Unsloth's 2-bit build, which the card warns may show quality differences in edge cases. No single consumer GPU comes close. The smallest build on Hugging Face, Unsloth's 1-bit UD-IQ1 S, is still 176 GB. How fast it runs on Macs Apple's machine learning team published its own test the same day. It linked Mac Studios with two Thunderbolt cables for RDMA, which lets one machine read the other's memory directly, using the jaccl backend in MLX v0.30.1 or later and macOS 26.2. The post says four Mac Studios generate over 36 tokens per second at an 8,192 token context, and its table lists 36.73. For two machines the table lists 15.74 tokens per second with pipeline parallelism and 24.99 with tensor parallelism. Treat the two-machine numbers with care: the post's tables merge two context-length columns, and one time-to-first-token figure in its text, 40.6 seconds, does not match the table. The command Apple shows for a local server is short: mlx.launch --hostfile hostfile.json --backend jaccl --env MLX METAL FAST SYNCH=1 -- mlx lm.server --model meta-models/Muse-Spark-Open-816B-A42B-NVFP4 --port 8080 Apple calls MLX METAL FAST SYNCH an experimental MLX feature. Where it wins and where it trails, from Meta's own table Meta's model card compares Muse Spark Open with seven other open-weight models at its highest reasoning setting, xhigh. All of these numbers are Meta's, and none of them come from an independent run. Where it leads: - MCP-Atlas, a tool-use test: 89.2, against 84.5 for Qwen3.8 and 84.2 for GLM 5.3. - JobBench, an agent task test: 61.8, just ahead of GLM 5.3 at 61.4. - HLE-Diamond with no tools: 26.1, against 22.2 for Kimi K3. - StrongREJECT v2, a jailbreak test where lower is better: 0.5, against 62.9 to 86.4 for the others. Where it trails: - The Artificial Analysis Intelligence Index: 39, behind GLM 5.3 at 45, Kimi K3 at 44 and Qwen3.8-2.4T-A95B at 40, and just ahead of Mistral Large 4 at 38. - Terminal-Bench 2.1: 76.4, against 85.0 for Kimi K3 and 83.9 for GLM 5.3. - DeepSWE 1.1, a coding test: 60, against 68 for Kimi K3 and 67 for GLM 5.3. - GPQA Diamond: 89.0, against 93.5 for Qwen3.8 and Kimi K3. That is why Meta's headline says most capable US open-weight model and not most capable open-weight model. On the overall index, three open models from other labs are ahead of it. Its strongest case is agents that call tools, plus a jailbreak resistance score far below any rival in the table. A day later, a second US lab released a rival On October 16, the day after Meta's release, Reflection released Beam, which Hugging Face's launch post calls the first model from the new US lab. It is the clearest comparison for anyone choosing an open model this week. - Size: 501B total parameters and about 23B active, against 816B and 42B for Muse Spark Open. - Context: 1M tokens for both. - 4-bit download: roughly 280 GB in NVFP4, about half of Muse Spark Open's 533 GB. - License: Apache 2.0, a standard open source license with no revenue line and no usage policy attached. Meta's own table already includes Beam. There Beam scores 80.1 on Terminal-Bench 2.1 against 76.4 for Muse Spark Open, and 90.5 on GPQA Diamond against 89.0. Muse Spark Open leads on MCP-Atlas, 89.2 to 78.7, and on SciCode, 58.5 to 49.7. Reflection reports 81.6% on SWE-Bench Verified for Beam, a test Meta's table does not include. So the trade is plain. Beam is smaller, cheaper to host and carries the simpler license. Muse Spark Open is bigger, reads images, video and audio as well as text, and leads on Meta's tool-use tests. What it costs through an API If you do not want to run it, OpenRouter lists it on October 16 at $1.25 per million input tokens and $4.25 per million output tokens, with cached input at $0.15 per million. All five providers it lists Modal, Together, Nebius, Fireworks and BaseTen charge that same price, and all five offer the full context window or close to it. That is also the exact price OpenRouter lists for Meta's closed Muse Spark 1.3. OpenRouter list prices for the rivals in Meta's table, per million input and output tokens, the same day: - Kimi K3: $0.30 in, $11.94 out. - Qwen3.8-2.4T-A95B: $2.00 in, $6.00 out. - Mistral Large 4: $0.68 in, $2.09 out. For a job that writes a lot, such as an agent that produces long code, the output price matters most, and Muse Spark Open is well under Kimi K3 there. Who is affected | Case | Status | |---|---| | Companies using it inside their own products or for internal work | Free at any size. The $20 million clause does not apply. You must show Built with Muse Spark Open and follow the usage policy. | | API resellers and cloud hosts with more than $20 million in total yearly revenue | Must notify Meta and sign a separate agreement before offering it as a service. | | Teams training smaller models on its outputs | Those models count as Derivatives, so the attribution rule and the usage policy carry over to them. | | Security teams and red teams | The usage policy bans creating malicious code, so exploit and malware testing sits in a gray zone. Read Exhibit A before you build on it. | | Teams that need the simplest license | Reflection's Beam, released a day later under Apache 2.0, is smaller about 280 GB at 4-bit and has no revenue line and no usage policy attached. | | Anyone with a single GPU or one ordinary workstation | Cannot run it locally. The smallest build is 176 GB. Use an API provider instead. | | Mac shops | Two 512GB M5 Ultra Mac Studios run the 4-bit build with MLX, per Apple's and Meta's guides. | | Products that could reach users under 18 | Meta says the model is not intended for minors, and the deployer must assess and reduce the risk. | What to do - Try it through an API first. OpenRouter's meta/muse-spark-open costs $1.25 in and $4.25 out per million tokens, the same at all five listed providers. - Before you ship it in a product, add the Built with Muse Spark Open line to your settings page or docs, and copy the license and usage policy into your distribution. - If your company sells AI access to others and has more than $20 million in yearly revenue, talk to Meta before you launch. - To self-host, size for the NVFP4 build 533 GB plus KV cache, about 254 KB per token at 16 bits by our estimate, or about 33 GB per 128,000 tokens of context. - Run your own tests on the tasks you care about. Every benchmark on the card is Meta's own run. - If you build security tools, read the malicious code clause in Exhibit A with your lawyer before you depend on the model. What is still unknown - Every benchmark score here comes from Meta's own model card. We found no independent run of the full table yet. - Our KV cache estimate assumes all 62 layers keep a full cache in 16-bit numbers. The card does not say whether any layers use a sliding window or a compressed cache, and an 8-bit cache would halve the figure. - Meta gives about 530 GB for NVFP4, Apple about 493GB, and the Hugging Face files add up to 533 GB. We do not know why Apple's figure is lower. - Apple's two-machine speed numbers come from tables that merge two context lengths, and one figure in its text does not match its table. - The license does not say whether a chatbot a company sells to clients counts as Model as a Service or as a feature of its own product. That line will matter to MSPs and agencies, and only Meta or a lawyer can settle it. - The model card lists FORTRESS scores as ARS 5.8 and ORS 17.4 without defining either. Muse Spark Open has the lowest ARS and the highest ORS in the table. - Meta says a full evaluation report is coming soon. It was not out on October 16. Sources AI News Report https://theainewsreport.com/ · every headline, every morning.