Meta releases Muse Spark Open, an 816B-parameter open-weight model built to run on Mac
2026-10-15
On October 15, 2026, Meta released Muse Spark Open, an 816B-total/42B-active-parameter mixture-of-experts model with a 1M-token context window accepting text, image, video, and audio input, aimed at coding and long-horizon agentic work. Per researcher Jack Rae, it is the first base model trained under Meta Superintelligence Labs, with additional training for the open release; several observers described it as roughly comparable to Meta's earlier Muse Spark 1.2. Weights ship under a custom Muse Model License requiring 'Built with Muse Spark Open' attribution and a separate commercial agreement for Model-as-a-Service businesses over $20M in annual revenue; researcher Nathan Lambert called the license 'trash.' Artificial Analysis scored it 39 on its Intelligence Index — first among US open-weight models but behind China's GLM-5.3 (45) and Kimi K3 (44) — with relative strength on the AA-Omniscience benchmark and weakness on AA's agentic evaluations, where GLM-5.3's entire lead over it is concentrated. Launch API pricing on Together/Fireworks/Baseten is $1.25/$4.25 per million input/output tokens. TechCrunch reported Meta partnered with Apple so a quantized (~530GB NVFP4) version runs locally on two 512GB Mac Studios via MLX, alongside support for 4x Blackwell Ultra or 8x H200 GPUs.
Significance 4: A frontier lab's new open-weight entrant topping a trusted independent intelligence index among US open models is a significant release, but it trails two Chinese open models and is explicitly weak on agentic evaluations, stopping short of changing the open-weights leaderboard picture.
Implications · machine-drafted, not owner judgment
This narrows the gap between Meta's open releases and the open-weights frontier currently held by Chinese labs (GLM-5.3, Kimi K3), but the agentic-evaluation weakness is notable given Meta's stated ambitions for long-horizon agents — it suggests the MSL reorganization hasn't yet closed the agentic-capability gap even where raw intelligence-index scores improve. The Mac-native deployment path, via Apple's MLX and a license built around a $20M MaaS-revenue threshold, targets prosumer/small-business local deployment rather than enterprise cloud, a different go-to-market from Meta's closed Wonder-line flagships.
- Whether a later MSL-trained flagship closes the agentic-evaluation gap Muse Spark Open shows against GLM-5.3.
- Independent (non-Meta) agentic benchmark runs, e.g. Terminal-Bench, for Muse Spark Open.
- Enforcement or disputes under the $20M MaaS-revenue licensing clause.
Sources
- curator [AINews] Agents disprove Polysemanticity; greatly improve Interpretability & Alignment retrieved 2026-10-16
- press Meta launches an open-weight AI model that runs on Macs retrieved 2026-10-16