Meta Releases Muse Spark Open at 816B With 42B Active and a 1M Context, but Local Run Needs About a Terabyte of Mac Memory

Meta released the weights for Muse Spark Open on October 15, 2026, calling it the most capable US open-weight model to date. The model is an 816 billion parameter mixture of experts with about 42 billion parameters active per token, a 1 million token context window, and native multimodal inputs for text, images, audio and video. Output is text only.

The open-weight pitch is real, and so is the hardware bill. Meta says the model runs on two Apple M5 Ultra Mac Studios with MLX. That setup needs about a terabyte of unified memory. Hosted APIs are live today on several partners, which is how most developers will actually try it.

We checked Meta’s research blog, the Hugging Face technical write-up by Pedro Cuenca, Carl Franzen’s VentureBeat report, and Apple’s MLX distributed inference post before writing this up.

What Meta shipped

Muse Spark Open sits in the Muse family alongside Muse Glimmer (the smaller dense open model) and the proprietary Muse Spark and Muse Wonder lines we covered when Meta launched Muse Wonder. Meta says it is built for coding and agentic work: long-horizon tasks, multi-file software engineering, and multi-step tool use.

Hugging Face’s architecture notes put the language model at 62 layers with a hidden size of 8192. Each MoE block has 256 experts and routes 8 experts per token. Checkpoints on Hugging Face are Muse-Spark-Open-816B-A42B (BF16) and Muse-Spark-Open-816B-A42B-NVFP4. The BF16 safetensors package is about 1.5 TB. A 1.8B perception encoder handles images and video frames. Audio is tokenized at 12.5 tokens per second.

Reasoning is always on. Effort levels are low, medium, high and xhigh. You can turn the effort down, but you cannot turn reasoning off.

Weights are listed on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models. The Hugging Face collection is at meta-models/muse-spark-open. API access is listed through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Supported engines include SGLang, vLLM, TensorRT-LLM, Transformers, Apple MLX, llama.cpp and Unsloth.

Meta also says custom Helion GPU kernels in SGLang deliver 1.03× to 3.95× kernel-level speedups on NVIDIA and AMD, plus a 3% end-to-end speedup on AMD MI350. Treat those as vendor figures until third-party serving benches catch up.

Benchmarks Meta published (and what is still loose)

Meta’s evaluation table, as carried by VentureBeat and Meta materials, puts Spark Open (xhigh) ahead of the listed rivals on JobBench and MCP-Atlas, but behind GLM 5.3 and Kimi K3 on several coding and professional-work rows. VentureBeat could not verify a separate public Artificial Analysis assessment of Spark Open at publication, so the Intelligence Index figure of 39 should stay attributed to Meta’s table.

Other Meta-reported xhigh scores include HLE-Diamond without tools at 26.1, GPQA Diamond at 89.0, OSWorld Verified at 80.4, and ZeroBench with tools at 50.0. DeepSWE rows use each model’s preferred harness, so the agent software is part of the comparison.

The geographic qualifier in Meta’s headline claim matters. “Most capable US open-weight model” is narrower than “best open-weight model worldwide.” Independent Artificial Analysis listings cited by VentureBeat put several Chinese open models above Meta’s reported Index of 39, including MiMo-V2.6-Pro at 46.

The Mac Studio reality check

Meta’s desktop story is two M5 Ultra Mac Studios acting like one machine with about a terabyte of unified memory via MLX. Apple’s ML research post measures NVFP4 Spark Open at an 8,192-token context: about 24.99 tokens per second on two M5 Ultras with tensor parallelism, and about 36.73 tokens per second on four.

Hugging Face separately tested llama.cpp Q2_K on a Mac M3 Ultra with 512 GB: about 18 tokens per second with the model using roughly 304 GB, and an oMLX oQ3e run at about 344 GB and about 21 tokens per second generation. Even the aggressive quants are not laptop toys.

Cost is the other catch. VentureBeat notes Apple lists a 256 GB M5 Ultra with a 1 TB SSD at USD 10,799. Four of those sum to about USD 43,196 for a terabyte of combined memory. That is reference math for memory capacity available today, not a Meta package price. Apple’s 512 GB Mac Studio configuration is due late October, with price still unannounced. Meta’s model card, as reported by VentureBeat, lists at least 700 GB of accelerator memory for NVFP4 and 2 TB for BF16.

License, gate and hosted pricing

Weights ship under the Muse Model License with limited commercial terms. Products must show “Built with Muse Spark Open” attribution. Model-as-a-service providers whose aggregate gross revenue with affiliates exceeds USD 20 million in any consecutive 12 months need a separate written agreement with Meta before offering the model or derivatives through that MaaS business. Internal use, research and non-MaaS product features are exempt under the wording VentureBeat and other secondary reports describe. Read the license yourself before you ship a hosted endpoint.

Hugging Face’s access gate asks for full legal name, date of birth and full organization name. VentureBeat notes the same pattern on other Meta model pages, so it is a precedent, not a one-off.

Together AI’s published serverless rates for Muse Spark Open are USD 1.25 per million input tokens, USD 0.15 per million cached input and USD 4.25 per million output. Those are Together’s rates, not a Meta-wide price. Together also ran a limited free promo through October 18, 2026 at 12:00 PM PT. Hosted Spark 1.x on Meta’s own API uses the same USD 1.25 / USD 4.25 standard rates for comparison, while Muse Wonder sits much higher at USD 4 / USD 20.

What it means for Indian developers

Most India-based teams will use a hosted API first. Together’s USD 1.25 / USD 4.25 rates match Meta’s own hosted Spark pricing band and undercut Muse Wonder by a wide margin. If you already budget against Spark 1.3, Spark Open on Together is in familiar territory, with the free promo window (if still open when you read this) worth burning for evals only.

Self-hosting is a different story. A terabyte-class Mac cluster or multi-GPU NVFP4 node is not a Bangalore coworking desk project. For local agents on modest hardware, Muse Glimmer under Apache 2.0 remains the practical Meta open option. Spark Open is for teams that need the big MoE and can pay for memory, or for teams that only need the weights for fine-tuning and distillation research under the license.

Watch the MaaS clause if you wrap Spark Open as a billed API for other companies. Internal tools and product features that are not model-as-a-service sit in the exemption language secondary reports describe, but Indian SaaS founders selling inference should get counsel on the USD 20 million affiliate-revenue trigger.

Meta is also running a parallel developer push. Applications for its USD 1M global AI hackathon close October 25, with India eligible. That credits pool is a cheap way to compare Meta’s hosted models against Spark Open endpoints before you commit infra.

How it sits against other frontier pricing

Open-weight does not automatically mean cheap to run. On the API side, Spark Open at Together lands near Meta’s hosted Spark and well below Claude Opus-class and Muse Wonder pricing. On the weight side, memory and serving cost dominate. Google’s Gemini 4 Argon intro pricing and OpenAI’s planned GPT-4o / GPT-4.1 API shutdown are separate calendars, but they are the practical alternatives many teams will price against while evaluating Meta’s open release.

Safety-wise, Meta says Spark Open met Meta Superintelligence Scaling Framework thresholds for release, with details in the model card and a fuller evaluation report promised soon. Until that report lands, treat the safety story as vendor-complete, not independently audited.

FAQ

What is Muse Spark Open?

It is Meta’s open-weight multimodal MoE released October 15, 2026: about 816B total parameters, about 42B active per token, up to 1M tokens of context, text and media in, text out, aimed at coding and agentic workflows.

Can I run Muse Spark Open on one Mac?

Not in the full configurations Meta and Apple highlight. Meta’s reference local setup is two M5 Ultra Mac Studios with MLX. Quantized llama.cpp and oMLX runs on a single 512 GB M3 Ultra are possible but still consume hundreds of gigabytes of RAM.

How much does the Together AI API cost?

Together lists USD 1.25 per million input tokens, USD 0.15 per million cached input and USD 4.25 per million output for serverless Muse Spark Open. Other hosts set their own rates.

Is the Muse Model License fully open for commercial SaaS?

Many uses are allowed, including internal and non-MaaS product use under the limited commercial terms Meta describes. MaaS businesses above USD 20 million in aggregate gross revenue over 12 months need a written Meta agreement. Attribution is required.

Are the benchmark wins independently confirmed?

Partly. JobBench, MCP-Atlas and other rows come from Meta’s table. VentureBeat could not verify a separate public Artificial Analysis score for Spark Open at publication, so keep the Index of 39 labeled as Meta-reported until an independent listing appears.