Meta's Muse Spark Open ships with 816 billion total parameters and 42 billion activated per token, a 1 million-token context window, and one unusual claim: it is the first model of its class that can run on a desktop. A dual-layer Blu-ray disc holds 50GB — fitting this model would take about twenty of them. The solution is two Apple Mac Studio machines, each with 512GB of unified memory, combining to 1TB. At BF16 precision, 816B parameters alone require roughly 1.6TB of weights, so the desktop-deployable version is quantized to NVFP4.
Key points
- Architecture and desktop deployment: Muse Spark Open is a mixture-of-experts model. Using the latest MLX framework and two M5 Ultra chips, tensor parallelism splits every layer in half — each Mac stores half the weights and computes half, with weights never moving between machines; only small intermediate results cross the wire at microsecond latency over Thunderbolt 5 with RDMA. The 1.2TB/s memory bandwidth per chip suffices to read the small slice of weights needed per token.
- Three deployment paths:
- On-device agentic demos: Meta's Muse Code, powered by Muse Spark Open, ran fully local workflows: submitting real iOS changes to the open-source Mastodon client Ice Cubes (bug fixes through complete builds), and end-to-end DaVinci Resolve control — reviewing footage frame by frame, selecting the best takes, and cropping to vertical format. Both are long, multi-step chains executed entirely on the machine. These are officially selected demos with no third-party reproduction; token throughput was not disclosed, so "it runs" is established, "it runs fast" is not.
- Benchmark positioning: Per aireleasetracker's transcription of Meta's scorecard (comparisons include Inkling, Nemotron 3 Ultra, Beam, Mistral 4 Large, Qwen3.8-2.4T-A95B, GLM 5.3, Kimi K3): FrontierCode v1.1 main set 53.3%, extended set 64.5%, Terminal-Bench 4.0 57.9%, SWE-Marathon v1.1 31.9%, FrontierSWE V2 25.3%. For context, Artificial Analysis measured Claude Opus 5.5 at 52% average on FrontierCode 1.1 with its own Stirrup harness — different harnesses mean the comparison only shows open weights are now in the same discussion range.
- Safety and licensing: Meta says open weights require extra care, so this release received additional safety post-training; evaluation covered the most severe risk categories in the Meta Superintelligence Scaling Framework, with a full report "to come." It uses the Muse Model License with limited commercial terms, free for many developers.
- Distribution breadth: Seven API providers (Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Together AI); four weight hosts (Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, Microsoft Foundry Models); inference engines including SGLang (with auto-tuned Helion GPU kernels, 1.03–3.95× kernel-level speedup, 3% end-to-end faster on AMD MI350), vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth; chip support spanning NVIDIA, AMD, Apple, and Intel.
| Configuration | Unified memory/VRAM | Weight format | Framework | Audience | |---|---|---|---|---| | 2× Mac Studio (M5 Ultra 512GB) | 1TB | NVFP4 | MLX | Desktops and personal studios | | 4× Mac Studio (256GB) | 1TB | NVFP4 | MLX | Without 512GB models | | 1 node, 4× Blackwell GPUs | 4 GPUs | NVFP4 | TensorRT-LLM + Dynamo | Scaling-out teams |
Positioning: second-best by design
Meta chief AI officer Alexandr Wang stated the tiering principle: the company will not open-source its strongest models, but will keep releasing somewhat lower-performance versions to developers. The release sits one tier below the closed Muse Wonder flagship published earlier that week — Meta itself acknowledges this. Zuckerberg's framing is more aggressive: freeing high-performance AI from cloud-server dependence.
Three takeaways for observers:
1. On-device sovereignty for coding agents is now a hardware spec competition — 1TB of memory is the new ticket to entry, out of reach for consumer Macs. 2. The "strongest American open weights" claim signals a shifted competitive landscape, with Qwen, GLM, and Kimi appearing in the comparison list. 3. Real ecosystem win: seven inference engines, four chip vendors, and auto-tuned kernels mean friction between "download weights" and "it runs" has been seriously addressed.
Unanswered questions matter just as much: actual local inference speed, the exact commercial boundaries of the Muse Model License, the timing of the full safety report, and when the 1TB barrier will drop. These determine whether this is developer productivity or an expensive demo.
Sources and caveats: Specs, deployment, demos, and licensing come from Meta's research and developer blogs (2026-10-15). Scorecard figures come from aireleasetracker's transcription and were not verified against the original chart; harnesses differ across evaluators. The Claude Opus 5.5 comparison is via Artificial Analysis's methodology. The Muse Wonder comparison and Wang's tiering principle come from press coverage. Local speed and license details were not disclosed at publication. Conversions (BF16 ≈ 1.6TB; 512GB×2 = 1TB) are computed from disclosed parameters.