# Muse Spark Open — What Changed, Why It Matters **Bottom line:** On **October 15, 2026**, Meta released the **open weights** for **Muse Spark Open**, an 816B-parameter sparse mixture-of-experts model it calls "the most capable US open-weight model to date." It's the open-weight version of the previously proprietary Muse Spark line, and it's notable for combining frontier-scale agentic/coding ability with the ability to run on a desktop. ## What changed **The release itself** - Meta released the **weights** for Muse Spark Open, a natively multimodal model with **816B total parameters (~42B active per token)** and a **1M-token context window**, built for coding and agentic work — long-horizon tasks, complex software engineering, and multi-step reasoning ([Meta AI Research](https://research.meta.ai/blog/muse-spark-open)). - It's the **open-weight version of Muse Spark**, which had been proprietary since its April debut; the starting point was **Muse Spark 1.2** (shipped August), with added safety post-training ([SiliconANGLE](https://siliconangle.com/2026/10/15/meta-releases-816b-parameter-muse-spark-open-for-coding-and-agent-work/)). - It follows **Muse Glimmer** (a much smaller ~30B dense model from August) — Muse Spark Open is the full-size model of the Muse series, sharing the same tokenizer/vocabulary ([Hugging Face blog](https://huggingface.co/blog/meta-muse-spark-open)). **Architecture & specs** (per [Hugging Face](https://huggingface.co/blog/meta-muse-spark-open)) - 62 layers, hidden dim 8192; **256 experts per MoE layer, 8 selected per token**; experts hold ~97% of parameters. - Inputs: **text, images, audio, and video**; output: text. Separate perception encoder (1.8B params) and an audio encoder (12.5 tokens/sec). - Checkpoints in **BF16 (~1.5 TB)** and **NVFP4** (4-bit, for NVIDIA Blackwell). BF16 runs across 4 nodes of 8×H200 (TP=32); NVFP4 runs on a single B200 node. **Availability** - **Hosted API:** Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Together AI. - **Weights download:** Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry Models ([Meta AI Research](https://research.meta.ai/blog/muse-spark-open); [HF collection](https://huggingface.co/collections/meta-models/muse-spark-open)). - **Inference engines:** SGLang, vLLM, TensorRT-LLM, HF Transformers, Apple MLX, llama.cpp, Unsloth — optimized for AMD, Apple, Intel, NVIDIA ([Meta dev blog](https://dev.meta.ai/resources/blog/build-with-muse-spark-open)). **License** - Released under the new **Muse Model License** with "limited commercial terms." No fee for developers who adapt it or ship it inside their own software, but a **separate agreement is required for companies selling it as a hosted service that have passed $20M gross revenue over 12 months**. Products must show a "Built with Muse Spark Open" label; **military and critical-infrastructure uses are off-limits** ([SiliconANGLE](https://siliconangle.com/2026/10/15/meta-releases-816b-parameter-muse-spark-open-for-coding-and-agent-work/)). ## Why it matters 1. **Frontier-scale open weights from a US lab.** Meta positions it as the most capable US open-weight model to date, and it's the first model in its class to run on a desktop — using **two Apple M5 Ultra Mac Studios** with MLX, whose unified memory (~1 TB combined) holds the whole model ([Meta AI Research](https://research.meta.ai/blog/muse-spark-open)). 2. **Strong agentic/tool-use results, mixed coding results.** In Meta's own benchmark table, Muse Spark Open scored **61.8 on JobBench** (vs. 31.1 for Thinking Machines' Inkling and 27.2 for Nemotron 3 Ultra) and **89.2 on MCP-Atlas**, beating every rival Meta listed. But on coding, Chinese open models led: **Kimi K3 topped Terminal-Bench 2.1 at 85**, with GLM 5.3 and Qwen3.8 also above Muse Spark Open's 76.4 ([SiliconANGLE](https://siliconangle.com/2026/10/15/meta-releases-816b-parameter-muse-spark-open-for-coding-and-agent-work/); [Unsloth](https://unsloth.ai/docs/models/muse-spark-open)). 3. **Real local/on-device agentic demos.** Meta ran Muse Code on the two-Mac setup to ship real iOS changes to the open-source Mastodon client **Ice Cubes** (bug fixes, features, full build) and to drive **DaVinci Resolve** end-to-end, reviewing footage frame-by-frame to crop vertical clips ([Meta AI Research](https://research.meta.ai/blog/muse-spark-open)). 4. **Practical local deployment is now feasible.** Unsloth shipped day-zero **Dynamic GGUFs**: the 2-bit quant fits on **256 GB** of RAM/unified memory, 4-bit needs ~512 GB, and 8-bit needs ~870–950 GB (full BF16 is ~1.63 TB). It supports low/medium/high/xhigh reasoning effort ([Unsloth](https://unsloth.ai/docs/models/muse-spark-open)). 5. **Safety framing.** Meta says the model has additional safety post-training for open release, was evaluated under its **Meta Superintelligence Scaling Framework**, and met the thresholds required for release — with a full evaluation report still to come ([Meta AI Research](https://research.meta.ai/blog/muse-spark-open)). ## Caveats - Several headline benchmark numbers (JobBench, MCP-Atlas, Terminal-Bench 2.1) come from **Meta's own benchmark table** as reported by third parties; the full independent evaluation report was not yet published at the time of writing. - Meta's **most capable systems remain closed** — the Muse Wonder models introduced days earlier are available only through the Meta Model API and Muse Code ([SiliconANGLE](https://siliconangle.com/2026/10/15/meta-releases-816b-parameter-muse-spark-open-for-coding-and-agent-work/)). ## Strongest sources - [Meta AI Research — Introducing Muse Spark Open](https://research.meta.ai/blog/muse-spark-open) (primary announcement) - [Hugging Face — Welcome Muse Spark Open](https://huggingface.co/blog/meta-muse-spark-open) (architecture & deployment detail) - [SiliconANGLE — Meta releases 816B-parameter Muse Spark Open](https://siliconangle.com/2026/10/15/meta-releases-816b-parameter-muse-spark-open-for-coding-and-agent-work/) (independent coverage, benchmarks, license) - [Unsloth — Run Locally Guide](https://unsloth.ai/docs/models/muse-spark-open) (quantization/hardware requirements) - [Meta Developer Blog — Build with Muse Spark Open](https://dev.meta.ai/resources/blog/build-with-muse-spark-open) (runtimes, partners, post-training) - [Hugging Face model collection](https://huggingface.co/collections/meta-models/muse-spark-open) (weights)