Muse Spark Open: Meta's Open-Weight Model for Coding and Agents, Explained

Muse Spark Open is Meta's new open-weight model for coding and agentic work, released October 15, 2026. It's the downloadable version of Muse Spark — an 816-billion-parameter mixture-of-experts model that activates about 42B parameters per token, reads up to 1M tokens of text, image, video, and audio, and returns text. Meta calls it the "most capable US open-weight model to date." Here's what that claim means, where the model leads, where it trails, what hardware it needs, and the license you actually agree to.

What Is Muse Spark Open?

Muse Spark Open is the open-weights release of Meta's Muse Spark line, built for long-horizon agentic work rather than one-shot chat. It plans, acts, observes, and self-corrects over long tasks — including browsing the web with images — and was trained on real developer workflows like multi-file refactors and long debugging sessions.

The headline specs from Meta's launch post:

• 816B total / 42B active parameters (a sparse mixture-of-experts design)
• 1M-token context window at inference
• Natively multimodal: text, image, video, and audio in; text out

Architecturally, it's a 62-layer transformer where each layer swaps a dense feed-forward network for a Mixture-of-Experts block of 256 experts, with a router picking 8 per token. A 1.8B-parameter perception encoder turns images and video frames into tokens, and a separate pathway handles audio. Because only 8 experts fire per token, the model runs at roughly the speed of a 42B model while needing the memory of an 816B one.

It is the open-weights version of Meta's Muse Spark model, with additional safety post-training so it holds up when run as standalone weights without system guardrails, per Meta's model card. Meta sells several closed models through its API — the frontier Muse Wonder models and Muse Spark 1.3 — while Muse Spark Open and the smaller Muse Glimmer are the open-weight releases you can download and self-host.

Is Muse Spark Open Actually Open Source?

No — and this distinction matters. The weights are downloadable, but they ship under Meta's custom Muse Model License, not an OSI-approved open-source license.

The license grants broad rights to use, modify, fine-tune, distill, and redistribute the model, but with conditions:

• Attribution: every product that uses the model or a derivative must prominently display "Built with Muse Spark Open" in its UI, docs, website, or settings.
• Model-as-a-service clause: a company offering API access to the model, with group revenue above $20M in any 12-month period, must notify Meta and sign a separate agreement first.
• Derivatives: anything fine-tuned or distilled from the model — even trained on its outputs — carries the same attribution duty.

For most teams — internal use, research, or features inside your own product — it's free to use commercially; the Muse Model License explicitly carves those cases out of the model-as-a-service clause. If you resell model access at scale, read the fine print first. The New Stack frames it bluntly: the weights are open but the license isn't fully open, a pattern it notes mirrors some recent Chinese releases.

Benchmarks: Where It Leads and Where It Trails

Meta positions Muse Spark Open as the best US open-weight model, and its own model card compares it at xhigh reasoning effort against seven similar open models, including Kimi K3, GLM 5.3, and Qwen3.8. The honest read is that it's a tool-use leader and a coding laggard.

Source: Meta's model card, October 2026.

Read the table by what your agent actually does:

• Tool selection: MCP-Atlas and JobBench are the rows where it leads, so agents whose main job is choosing and sequencing tool calls are its strongest fit.
• Terminal and repo work: it trails Kimi K3 by 8.6 points on Terminal-Bench 2.1 and by 8 points on both DeepSWE 1.1 and SWE-Atlas-QnA. If your agent lives in a terminal, run your own repository tasks before switching.
• Composite rank: on the Artificial Analysis Intelligence Index it scores 39, behind GLM 5.3, Kimi K3, and Qwen3.8 — so "most capable" is specifically a ranking among US open-weight models, not the whole open field.

One area where it genuinely stands out is jailbreak resistance: it scored 0.5 on StrongREJECT v2 with adaptive attacks (lower is better), where no other model in Meta's table scored below 62.9 — a direct payoff of that extra safety post-training.

What Hardware Does Muse Spark Open Need?

This is the catch. "Open weight" does not mean "runs on your laptop." The full BF16 checkpoint is about 1.6 TB, and even the quantized NVFP4 build is around 530 GB.

• GPU: NVFP4 serves on four Blackwell Ultra GPUs or eight H200s; the model card asks for at least 700 GB of VRAM for NVFP4 and 2 TB for BF16.
• Mac: Meta and Apple tuned an MLX build that runs the NVFP4 checkpoint on two M5 Ultra Mac Studios with 512 GB of unified memory each, or four Mac Studios with 256 GB each — roughly 1 TB of unified memory total — linked over Thunderbolt 5.
• Smallest build: Unsloth's GGUF quants go down to about 2-bit and fit in roughly 300 GB of combined memory — a single high-memory workstation, not a MacBook.

It runs across SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp, and Unsloth, and is available day one on hosted APIs including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Weights download from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. For most teams, a hosted endpoint is the realistic way to try it before committing to hardware.

Who Should Use It?

Muse Spark Open is a strong pick if you're building agents that orchestrate tools — the MCP-Atlas and JobBench results back that up — and you want the option to self-host for data control, fine-tune on your own workflows, or distill a smaller model. The native multimodality is a real differentiator: the same model can read a screenshot, listen to a voice note, or watch a screen recording while it works.

It's a weaker pick if your agent is mostly a coding agent living in a terminal, where open rivals like Kimi K3 currently score higher. And it's not for anyone expecting to run frontier-scale weights on consumer hardware.

Bottom Line

Muse Spark Open is a credible, genuinely useful open-weight release that's best understood as a tool-use and multimodal agent model, not a coding champion. The "most capable US open-weight model" framing is accurate but narrow. The license is permissive for most teams but worth reading if you resell access, and the hardware bar is high enough that a hosted API is the sensible on-ramp. If you build agents and value self-hosting, it belongs on your shortlist to test against your own tasks.

Run Open Models Inside Your Own AI Workforce

An open-weight model earns its keep once it's wired into real workflows — tools, files, code review, and multi-step plans — not a chat window. Eigent is an open-source Cowork desktop app that runs a multi-agent AI workforce locally and is model-agnostic, so you can route a model like Muse Spark Open to the tasks it's strongest at and keep sensitive work on your own machine. See how agents can review GitHub PRs end to end, then download Eigent to try it.

Frequently Asked Questions

What is Muse Spark Open?

Muse Spark Open is Meta's open-weight model for coding and agentic work, released October 15, 2026. It's an 816B-parameter mixture-of-experts model that activates about 42B parameters per token, accepts text, image, video, and audio input, and has a 1M-token context window. It's the open-weights version of Meta's Muse Spark model.

Is Muse Spark Open open source?

No. The weights are downloadable, but they're released under Meta's custom Muse Model License, not an OSI-approved open-source license. The license allows use, fine-tuning, distillation, and redistribution, but requires a visible "Built with Muse Spark Open" label in every product that uses the model or a derivative.

Is Muse Spark Open free for commercial use?

Yes, for most teams. Only a company offering models as a service through an API, with group revenue above $20M in any 12-month period, must notify Meta and sign a separate agreement first. Internal use, research, and features inside your own product fall outside that clause.

Can Muse Spark Open run on a Mac?

Yes, but you need serious hardware. The NVFP4 checkpoint runs through MLX on two M5 Ultra Mac Studios with 512 GB of unified memory each, or four Mac Studios with 256 GB each, linked over Thunderbolt 5 — roughly 1 TB of unified memory in total. Unsloth's smallest GGUF build fits in about 300 GB of combined memory.

Is Muse Spark Open the best open-weight model for coding?

No. On Meta's own model card it scores 76.4 on Terminal-Bench 2.1, 60 on DeepSWE 1.1, and 58 on SWE-Atlas-QnA — behind Kimi K3 on all three. Its lead is in tool use, where it tops MCP-Atlas at 89.2. Run your own repository tasks before moving a coding agent onto it.

How does Muse Spark Open relate to Meta's closed Muse models?

Muse Spark Open is the open-weights version of Meta's Muse Spark model, with additional safety post-training so it resists misuse when run without system guardrails. Meta also sells closed models through its API — the frontier Muse Wonder models and Muse Spark 1.3 — while Muse Spark Open's weights download from Hugging Face so you can self-host, quantize, and fine-tune it.