Meta announced Muse Spark Open, billed as the “most powerful open-weight model from the U.S. to date”: ~816B total parameters (MoE, ~42B active), ~1M-token context, native text/image/video/audio, aimed at long-horizon coding and multi-step reasoning. Per Meta and reports, two M5 Ultra Mac Studio machines (~1TB unified memory combined) can run it locally; weights land on Hugging Face and major cloud/API hosts the same day.
Specs that matter
| Item | Per Meta / reporting | |——|———————-| | Params | ~816B total / ~42B active (MoE) | | Context | ~1M tokens | | Modalities | Native text, image, video, audio | | Focus | Long tasks, hard software engineering, multi-step reasoning | | License | Muse Model License — run/modify/fine-tune/distill/redistribute with conditions; if licensee + affiliates exceed ~$20M trailing-12-month revenue and offer the model (or derivatives) as a service, notify Meta and sign a separate written deal first |
Distribution: Hugging Face plus AWS SageMaker, Google Cloud, Microsoft Foundry; Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, Together AI and peers reportedly day-one.
For prior Muse / Meta agent coverage on this site, see Muse going viral on the CPU side and Meta’s shift toward “personal superintelligence”.
Why “two Mac Studios”
A maxed M5 Ultra Mac Studio is said to offer up to ~512GB unified memory; two linked units ≈ 1TB, with ~1.2 TB/s bandwidth, under Apple MLX. Meta AI lead Alexandr Wang argued many strong open models are impractical on ordinary PCs, while Mac Studio memory makes Muse Spark Open “practical.”
Cost: ~$10,799 per loaded studio, ~$20k+ for a pair. On a four-studio cluster, throughput is cited above ~36 tok/s. Beyond Apple, Meta claims NVIDIA/AMD/Intel plus SGLang, vLLM, TensorRT-LLM, llama.cpp; on four NVIDIA Blackwell GPUs with speculative decoding, reports cite peaks above ~50k tok/s per GPU and ~600 tok/s per user (verify against official benches).
On-device agent demos
Meta showed Muse Code workflows: shipping real iOS changes to the open Mastodon client Ice Cubes (bugs, features, full builds), and driving DaVinci Resolve end-to-end—frame review, highlight picks, vertical crops. The pitch is offline, on-device agent labor—not another metered cloud API.
Meta × Apple: rivals, shared hardware story
Public posts had Mark Zuckerberg thanking Apple CEO John Ternus for Mac Studio performance; Ternus congratulated the team on pushing on-device AI. Context: ATT-era privacy fights once cost Meta on the order of ~$10B ad revenue (2022 estimate), while cloud-scale AI still sits with hyperscalers—Meta wants terminals that can host giant open weights; Apple wants Mac restored as the top AI workstation narrative.
FAQ
Q: How does an 816B model fit on two desktops? A: MoE (~42B active) plus Apple unified memory (~1TB across two boxes) and MLX—not a free pass for arbitrary PCs.
Q: Is commercial use free? A: Broad rights under Muse Model License for most builders; crossing ~$20M TTM revenue while offering the model/derivatives as a service triggers Meta notice + a separate agreement.
Q: Why not just use cloud APIs? A: Local run, no per-token meter, data stays on-device—paid for up front in hardware and ops.
相关文章
