Muse Spark Open decoded: Meta's 816B open-weight coding model, hosted or on Macs

Meta's 816B open-weight model is on seven API hosts and two-Mac setups; for builders routing coding agents to open models and weighing self-hosting.

Picture a four-person agency that builds internal tools for clients. Their coding agent routes to an open-weight model through a hosted API because it's cheaper than a closed frontier model. Then a client in a regulated industry asks that its code never leave hardware the agency controls. The team now has two bad options. They can rent a GPU node they don't know how to run, or they can drop to a small local model that falls apart on multi-file refactors.

Meta's Muse Spark Open, released October 15, 2026, aims at that gap. It is an open-weight model built for coding and agents, available on seven API hosts on day one. Meta and Apple also say it runs on two Mac Studios with M5 Ultra chips. The real questions are whether it beats the open models you already route to, and whether "run it at home" means much when home is a terabyte of unified memory.

What it is

According to Meta's announcement, Muse Spark Open is a natively multimodal mixture-of-experts model:

Hugging Face's write-up adds that it takes text, image, audio and video input and outputs text. It has 256 experts per layer, with 8 picked per token, and ships in BF16 and NVFP4 checkpoints.

Here is how each piece maps to the agency's problem.

On quality, Artificial Analysis scores it 39 on its Intelligence Index at xhigh reasoning, against a median of 19 for comparable open models. It measures a median of 130.7 tokens per second and 0.73s to first token across providers. On X, Artificial Analysis calls it the top-scoring US open-weights model, ahead of NVIDIA's Nemotron 3.5 Super VL at 26. The Decoder and TechCrunch both note that the claims of beating Qwen, GLM and Kimi come from Meta's own evaluations.

Alternatives, as the sources describe them. We didn't fetch these models' own pages.

Who it's for

Pick it up if one of these fits:

Skip it if one of these fits:

Before and after

Before (the agency today):

  1. Route agent tasks to a hosted open model.
  2. For the privacy-bound client, either rent and run a GPU node, or use a small local model and accept weaker results on long tasks.
  3. Maintain two prompt setups, one per model.

After (with Muse Spark Open):

  1. Point the agent at Muse Spark Open on a hosted provider through OpenRouter or a direct host.
  2. For the private client, serve the same weights locally. That means two 512GB M5 Ultra Mac Studios with MLX, or four 256GB ones, per Meta's developer post, or one four-GPU Blackwell node.
  3. Keep one model and one prompt setup across both deployments.

The catch is in step 2's speed. In Apple's tables as extracted by explainx.ai, decode speed varies by setup:

| Setup | Tokens per second | |---|---| | Two M5 Ultras, pipeline | 15.74 | | Two M5 Ultras, tensor parallel | 24.99 | | Four M5 Ultras | 36.73 |

All of these are at batch size 1, which means one user. explainx.ai also flags that these are Apple-run benchmarks.

Try it

The cheapest first step is hosted. Run your own prompts through OpenRouter's listing, which shows several providers at $1.25/$4.25 with different latencies. TokenCost reports that Together runs a separate promo endpoint at $0 until October 18, so that window is about to close.

To run it locally with llama.cpp, Meta's developer post gives this command for a 4-bit GGUF:

llama serve -hf ggml-org/Muse-Spark-Open-816B-A42B-GGUF:Q4_K_S

Before you run it, check two things.

For Macs, Meta points to an MLX adapter and a "Muse Spark Open on Mac" cookbook. The sources don't reproduce the MLX launch steps, so follow that cookbook rather than guessing.

The hardware cost is real. CNET reports the M5 Ultra Mac Studio starts at $5,499 with 96GB, that 256GB of RAM adds $4,000, and that the 512GB configuration isn't on sale yet.

Where it breaks

The honest verdict: worth evaluating if you need large open weights you can run on your own hardware, multimodal input, or a Mac-based private setup. Hard to justify as a rented API for agent-heavy coding when the sources show GLM-5.3 scoring higher on coding for a little more money.

Sources