채팅 구독만 말고, 최강 오픈웨이트를 내 책상에

Sources — Muse Spark Open (2026-10-16)

Official (primary)

·Meta AI Research blog (2026-10-15): https://research.meta.ai/blog/muse-spark-open

·Released weights Oct 15, 2026

·816B total / 42B active, 1M-token context, natively multimodal

·Built for coding and agentic / long-horizon / multi-step reasoning

·Desktop: two Apple M5 Ultra Mac Studios + latest MLX (first in class claimed)

·Hosted APIs: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Together AI

·Weights: Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, Microsoft Foundry Models

·Inference engines: SGLang, vLLM, TensorRT-LLM, HF Transformers, Apple MLX, llama.cpp, Unsloth

·On-device demos claimed: Muse Code → Ice Cubes iOS changes; DaVinci Resolve end-to-end

·Helion kernels: 1.03–3.95× kernel-level on NVIDIA/AMD (blog); not used as primary slide claim

·License: Muse Model License with limited commercial terms; many developers free (do not invent commercial details)

Secondary

·Hugging Face collection: https://huggingface.co/collections/meta-models/muse-spark-open

Uncertainties / careful wording

·"Most capable US open-weight" is Meta's own claim — slides say "미국 오픈웨이트 중 최강급이라며" (attributed).

·Commercial license details beyond "many developers free / limited commercial terms" were not invent­ed; caption points readers to the license text.

·Local desktop requires two M5 Ultra Mac Studios — not a typical laptop setup; slides emphasize API-first for most readers.