Introducing Muse Spark Open: The Most Capable US Open-Weight Model to Date

On Meta's own numbers, Muse Spark Open wins the American open-weight race by a wide margin and still trails in the global one.

Meta released Muse Spark Open on Thursday: an 816-billion-parameter, natively multimodal mixture-of-experts model with a one-million-token context window, built for coding and long-horizon agentic work. Meta calls it the most capable open-weight model from a US lab to date, and against its US peers the benchmarks back that up. Against the best Chinese open models the picture is more mixed: it is ahead on MCP-style tool use and visual reasoning, and behind on terminal-driven coding.

That split is the real story, so this post follows it: what the model is, where the evidence backs Meta up, where it doesn't, and what it takes to run.

The short version

Why this release matters

Open weights are the go-to option for teams that need control over cost, latency, and where their data lives. Chinese labs currently hold much of that ground, and Meta has said it believes the US should not let Chinese labs lead on open-source AI. Muse Spark Open is its answer. Among US open models, the nearest peers are Thinking Machines' Inkling and Nvidia's Nemotron 3 Ultra, and in Meta's own benchmarks Muse Spark Open beats both across the board, often by 20 points or more.

What's under the hood

Sparse by design. Each token is routed to a small subset of experts, so per-token compute tracks the 42B active parameters rather than the full 816B. The catch is that serving still loads every weight into memory, and that is what sets your hardware bill. The model card lists 62 layers, a hidden dimension of 8,192, and 256 experts per MoE layer with 8 active per token.

Perception built in. The model reads text, images, audio and video natively. Meta says its visual reasoning runs through a real execution environment instead of scripted steps, which matters for agents that browse the web with images in the loop. A separate 1.8B-parameter perception encoder handles visual input.

Hardened for the open. Muse Spark Open is derived from Meta's Muse Spark, with additional safety post-training meant to make it more robust against misuse as a standalone model. That is a sensible call for open weights: once the checkpoint is on your disk, there are no hosted guardrails to lean on.

An ecosystem that shipped with it. Meta released Helion kernels for operations like RoPE, RMSNorm and quantization, reporting 1.03x to 3.95x speedups over custom kernels written in hardware-specific languages. Inference is tuned for SGLang and vLLM across NVIDIA, AMD and Intel hardware, and TorchTitan supports supervised fine-tuning and reinforcement learning, so you can adapt the model rather than only prompt it.

Adjustable reasoning. Reasoning depth is set with a reasoning_effort parameter (low, medium, high or xhigh). The default is xhigh, so expect your own cost and latency to vary with the setting.

Built for agents, not just autocomplete

Meta trained Muse Spark Open on real developer workflows: multi-file refactors, long debugging sessions, reading a codebase and calling tools along the way. The intended loop is plan, act, observe, self-correct, sustained over long tasks.

Meta's showcase paired the model with its Muse Code terminal and used it to ship updates to the Ice Cubes Mastodon client, covering bug fixes, new features and a full build. Meta also pitches it for orchestrating multi-agent workflows with a single model and for customer-service agents.

The benchmarks, read honestly

Every number below comes from Meta's published comparison, run at maximum reasoning effort, and some rival scores are self-reported by their developers. Treat them as vendor-reported until independent results land.

[embedded content: Meta's published comparison as reported by MarkTechPost, Oct 15, 2026 · rival = the Chinese open model cited per benchmark; only Kimi K3 was cited for ZeroBench and HLE-Diamond]

%%{init: {"themeVariables": {"xyChart": {"plotColorPalette": "#2b7bd6"}}}}%%
xychart-beta
    title "Where Muse Spark Open leads: points ahead of the rival cited"
    x-axis ["ZeroBench", "MCP-Atlas", "HLE-Diamond", "JobBench"]
    y-axis "Points ahead" 0 --> 10
    bar [9.0, 4.7, 3.9, 0.4]

%%{init: {"themeVariables": {"xyChart": {"plotColorPalette": "#9a9a90"}}}}%%
xychart-beta
    title "Where it trails: points behind the rival cited"
    x-axis ["Terminal-Bench", "DeepSWE", "AA Index", "Toolathlon", "HLE+tools"]
    y-axis "Points behind" 0 --> 10
    bar [8.6, 7.5, 6.0, 3.7, 1.5]

The table below lists the underlying scores and which rival each gap is measured against.

The lead is not uniform: Kimi K3 and GLM 5.3 edge it out on Toolathlon Verified, and Muse Spark Open roughly ties them on SciCode. A weak coding score is a notable tension for a model pitched first and foremost for coding, and it echoes a gap Meta already acknowledged with the closed Muse Spark. For MCP-heavy automation, multi-step business workflows or image-grounded agents, it is a leading open option on Meta's numbers; for hard software-engineering benchmarks, the Chinese models still score higher.

Scale matters in this comparison. Kimi K3 (2.8T total, 104B active) and Qwen3.8 (2.4T total, 95B active) are far larger, while GLM 5.3 (753B total, roughly 40B active) is close in size. Muse Spark Open is also the only model in that group that accepts audio input.

Running it: the hardware reality

It's open, but "run it yourself" comes with a price tag.

• BF16: the full-precision reference weighs roughly 1.6 TB and serves on eight GPUs, such as 8 x B300.

• NVFP4: the quantized build covers only the experts and lands near 530 GB. Meta's guide lists 4 Blackwell Ultra GPUs or 8 H200s, and the model card recommends at least 700 GB of VRAM for NVFP4 and 2 TB for BF16.

• Hosted: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI serve it, so you don't have to provision GPUs. OpenRouter lists $1.25 per million input tokens and $4.25 per million output tokens.

• On a Mac: Muse Spark Open was tuned with Apple for Apple silicon through MLX. "Local" is relative, though: you need roughly 1 TB of unified memory, which means two linked 512 GB M5 Ultra Mac Studios or four 256 GB machines (about $38,000, per Engadget). Apple has said the 512 GB model arrives in late October without announcing pricing.

• On a workstation: Unsloth's 2-bit GGUF builds fit in about 300 GB of combined memory.

The fine print: open weights, not open source

The weights are downloadable, and the Muse Model License permits use, fine-tuning, distillation and redistribution, but it is not a permissive open-source license, and it's worth reading before you build a business on it:

• Revenue threshold. A company with more than $20 million in gross revenue in any consecutive twelve months must notify Meta and sign a separate written agreement before offering the model or its derivatives as a Model-as-a-Service business.

• Attribution. Products or services using the model, or anything trained, fine-tuned or distilled from it, must prominently display "Built with Muse Spark Open" in the interface.

• Free where it counts. Internal use and building the model into your own product are free.

• Use restrictions. The accompanying usage policy bars military and critical-infrastructure deployments.

The terms look much like the license Moonshot uses for Kimi K3. I'm not a lawyer, so if you plan to resell inference or sit near the revenue threshold, read the full text and get proper advice.

Who should use it, and who should wait

Use it if you are building agents or tool-calling pipelines that must keep prompts and responses inside your own infrastructure, want to fine-tune with SFT or RL on a large base, or prefer a US-developed open model for procurement or policy reasons. Look elsewhere if your work is dominated by terminal-driven coding, you can't fund the hardware or a hosted provider, or you plan to resell the model above the revenue threshold.

Getting started

Weights live on Hugging Face under meta-models/Muse-Spark-Open-816B-A42B, with an NVFP4 variant alongside. A minimal vLLM launch looks like this:

vllm serve meta-models/Muse-Spark-Open-816B-A42B-NVFP4 \
  --max-model-len 1048576 \
  --reasoning-parser muse \
  --tool-call-parser muse

Check the model card for parallelism settings and Meta's recommended speculative-decoding configuration before pointing production traffic at it.

The bottom line

Is Muse Spark Open the most capable US open-weight model? On Meta's numbers, comfortably. Is it the best open model overall? No: it wins on MCP-style tool use and visual reasoning, and trails on terminal-style software engineering and on the aggregate Artificial Analysis index. The license has strings and the hardware bill is real, but for anyone building agents on open weights, the field just got a serious new contender. Keep an eye out for independent evaluations, and for Apple's 512 GB Mac Studio later this month.

Sources

• Meta, Muse Spark 1.2 Is Now Open Weights (PyTorch blog)

• Meta, Muse Spark Open model page and documentation

• The New Stack, Meta ships Muse Spark Open, its answer to China's open-weight models

• Hugging Face, Welcome Muse Spark Open

• Engadget, Meta built a huge open-weight AI model for Macs, but you'll need 1TB of RAM to run it

• TechCrunch, Meta launches an open-weight AI model that runs on Macs

• MarkTechPost, Meta Releases Muse Spark Open