Meta Muse Spark Open Review: 816B Open Weights, Leading Agent Tool Use

The weights are on Hugging Face as meta-models/Muse-Spark-Open-816B-A42B, with downloads requiring manual approval, in BF16 and NVFP4 precision. The publisher is Meta Superintelligence Labs; the release date is October 15 under Meta's custom Muse Model License. Day-one hosted APIs came from Together AI, Fireworks AI, Baseten, Groq, Modal, Nebius, and OpenRouter. The same day Meta announced a partnership with Apple: the model is optimized for Apple's MLX framework and can be split across two M5 Ultra Mac Studios. Mark Zuckerberg posted on X that the weights were releasing that day; Apple CEO John Ternus replied publicly, calling it a showcase for on-device AI.

Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple.

— Mark Zuckerberg

Congrats to the team! Amazing to run a model of Muse Spark Open's caliber and scale this efficiently, right from your desk. Mac Studio keeps pushing the boundaries of on-device AI!

— John Ternus

This review did not run the model locally: the full 816B weights need terabytes of VRAM, beyond this machine. Testing had two parts. First, hands-on verification of official channels: the official Hugging Face repo exists, downloads require manual approval, it holds 65 safetensors weight shards (78 files total), and the page marks the license as a custom license (not an OSI open-source license); OpenRouter's live model list sells meta/muse-spark-open at $1.25 per million input tokens and $4.25 per million output tokens with 1,048,576 tokens of context, matching the official figures. Second, published verifiable scores: Meta's official model card (2026-10-15) reports 89.2 on MCP-Atlas tool use, 61.8 on JobBench, 76.4 on Terminal-Bench 2.1, 60 on DeepSWE, 25 on AA Omniscience (hallucination resistance), and 0.5 on StrongREJECT with 99.2% of jailbreak attempts refused — all Meta's own test data, not independently verified. The independent Artificial Analysis Intelligence Index v4.3.2 (10 evaluations weighted, tested at xhigh reasoning effort) scores it 39, the highest US open-weight score, 10th overall; the same page measures 141 tokens/sec output and 0.80s time to first token.

1. Architecture: Sparse mixture-of-experts: 816B total parameters, 42B active per token, 256 experts with 8 routed per token, 62 layers, 202,048 vocabulary; text, image, audio, and video input, text-only output; 1M-token context; four reasoning effort levels, xhigh by default.
2. License: Muse Model License, a custom license: products must display “Built with Muse”; API businesses with over $20M annual revenue must notify Meta and sign a separate agreement first; fine-tuning, distillation, and commercial use are allowed; Hugging Face downloads require name, date of birth, and organization.
3. Hardware bar: The full BF16 release is about a 1.65TB download needing ~2TB of VRAM; the NVFP4 release is about 537GB needing at least 700GB of VRAM; the community GGUF quantized build needs ~300GB of combined RAM and VRAM; the Apple MLX build needs two M5 Ultra Mac Studios with 512GB unified memory each linked over Thunderbolt 5 — that 512GB configuration had not shipped as of publication.
4. Vs. Muse Wonder: Wonder, released three days earlier, is a closed flagship available only via API at $4/$20 per million tokens; Spark Open is the downloadable open-weight version, priced at $1.25/$4.25 per million tokens via API.
5. Third-party take: Independent researcher Nathan Lambert called it the strongest American-made open model “by quite some margin” and said Meta is the only American company with frontier open and closed models; Chinese labs previously led that leaderboard. (via theaitools.tech)