Meta released the weights for Muse Spark Open on October 15 — 816 billion parameters, 42 billion active, a one-million-token context window, and multimodal inputs covering text, image, video, and audio. It is the most capable US-built open-weight model released to date. It also trails Kimi K3 by nine points on coding benchmarks. Both facts are true, and both matter if you are deciding whether to use it.
The Hardware Reality Nobody Wants to Say Out Loud
Before benchmarks, before pricing: most developers cannot run this locally. The BF16 weights come in at 1.6 TB. Meta’s recommended NVFP4 format needs roughly 530 GB of VRAM, which means four Blackwell Ultra GPUs or eight H200s. The “it runs on Mac” headlines refer to two M5 Ultra Mac Studios with 512 GB of unified memory each, linked over Thunderbolt 5 — a setup that costs well over $15,000 before you account for the Thunderbolt bridge.
Unsloth’s 2-bit GGUF quantization brings the floor down to around 300 GB. That is still not a single consumer GPU. The honest framing: Muse Spark Open is a model you run on cloud infrastructure or in an enterprise GPU cluster, not on your workstation. The Mac story is real but narrow.
Hardware tiers at a glance:
- Cloud API (no hardware): Together AI, Fireworks, Baseten, Modal, Nebius — available now
- Apple Silicon: 2x Mac Studio M5 Ultra (512 GB each) via MLX — ~1 TB unified memory total
- GPU with GGUF: Unsloth GGUF quants — Q4_K_S at ~470 GB, or 2-bit at ~300 GB
- Full GPU cluster: NVFP4 at ~530 GB across 4x Blackwell Ultra or 8x H200
Where Muse Spark Open Actually Wins
The coding benchmark story is mixed. On Terminal-Bench 2.1, Muse Spark Open scores 76.4; Kimi K3 scores 85.0 and GLM 5.3 scores 83.9. On DeepSWE 1.1, the gap is 60 versus 68. Meta publishes these numbers in its own model card, which takes some integrity. Coding is the primary reason most developers self-host large models, so this is not a footnote — it is the main story.
What the model genuinely excels at is agent workflows. On MCP-Atlas, the tool-use benchmark, Muse Spark Open scores 89.2 — top of Meta’s comparison table, ahead of every Chinese model listed. On JobBench it scores 61.8, beating Kimi K3’s 54.3 by a meaningful margin. On HLE-Diamond reasoning without tools, it scores 26.1 against Kimi K3’s 22.2. The pattern: if your use case involves multi-step agent orchestration, tool calling, or complex reasoning — rather than write-and-run code generation — Muse Spark Open holds up well.
The architecture supports this. It is a Mixture-of-Experts Transformer paired with a Perception Encoder: 62 layers, 256 experts, 8 activated per token via top-k routing. Supported inference frameworks include SGLang, vLLM, Hugging Face Transformers, TensorRT-LLM, Unsloth, and Llama.cpp. Full details are in Meta’s developer guide.
Cloud Inference Is the Real Starting Point
For teams that want to use the model without owning a GPU cluster, cloud inference is available now. Together AI, Fireworks, Baseten, Modal, and Nebius all serve Muse Spark Open at a uniform rate: $1.25 per million input tokens, $4.25 per million output tokens, and $0.15 per million cached tokens. The pricing is uniform because the Muse Model License gives Meta explicit control over what large hosts charge. Groq is serving it in limited preview at roughly 350 tokens per second with a 131K context window. OpenRouter aggregates all available providers if you want to route across them.
The License — Read Before You Ship
Muse Spark Open is released under the Muse Model License, not Apache 2.0. What is permitted without restriction: internal use, research, fine-tuning, distillation, derivative models, and building features inside your own products. What requires a separate agreement: selling it as a Model-as-a-Service once you and your affiliates exceed $20 million in gross revenue in any consecutive 12 months. Most teams building products will never hit that threshold.
Attribution is mandatory. You must display “Built with Muse Spark Open” prominently — in the UI, documentation, or settings. Prohibited uses include critical infrastructure, transportation systems, military applications, espionage, and nuclear industries. The license is restrictive at the edges but workable for the vast majority of commercial product development. A detailed breakdown is available at importstatic.com.
Who Should Use This
US-based enterprises with data residency requirements or geopolitical constraints on Chinese-developed models now have a genuine frontier-class open-weight option. That is the honest niche Muse Spark Open fills — and it is a real one. Teams building agent pipelines where tool use and multi-step reasoning matter more than raw coding speed will find it competitive. Teams whose primary workload is code generation should benchmark it carefully against Kimi K3 before committing infrastructure.
For the rest: the cloud API costs less than running your own inference cluster, the model card is more transparent than most, and the weights are downloadable from Hugging Face. This is a meaningful release from Meta — just not the one that closes the gap with Chinese rivals on every benchmark.

