Muse Spark Open
muse-spark-open- - 🏢 Meta's open-weight reasoning model for coding and agentic workflows.
- - 🧠 Sparse mixture-of-experts: 816B total, roughly 42B active per token.
- - 📏 One-million-token context window for whole-repository work.
- - 👁️ Native image understanding alongside text in a single checkpoint.
- - 🔧 Tool calling supported for agentic, multi-step developer loops.
- - 🆕 Open-weight counterpart to Meta's hosted Muse Spark line.
- - 🌐 Distributed on Hugging Face as a downloadable checkpoint.
- - 💬 Produces explicit reasoning before final answers.
Meta Platforms is an American multinational technology company headquartered in Menlo Park, California. Best known for operating Facebook, Instagram, WhatsApp, and Threads, Meta has become one of the most influential players in open-weight AI research through its Llama family of…
Explore 6 more models by Meta Platforms →Muse Spark Open is Meta's open-weight reasoning model aimed at coding and agentic workflows, released in 2026 as the downloadable sibling to the hosted Muse Spark models served through Meta's Model API. Architecturally it is a sparse mixture-of-experts network with roughly 816 billion total parameters and about 42 billion active per token, combining text generation with native image understanding in a single checkpoint and supporting tool calling and a context window of one million tokens.
Within Meta's own open-weight line, the clearest step is scale and modality. Hugging Face's launch write-up notes that Muse Spark Open follows Muse Glimmer, a far smaller open release, by introducing the mixture-of-experts architecture "at a much larger scale." Compared with Meta's earlier dense, text-only open models such as Llama 3.3 70B, the smaller Llama 3.2 3B, and the community-tuned Hermes 3 Llama 3.1 405b, Spark Open adds native visual input, explicit reasoning traces, tool calling, and a context window orders of magnitude longer.
The design is agent-oriented rather than purely conversational: the model is intended to plan, call tools, read results and continue over long horizons, which is what the very long context and function-calling support are there to enable.
For hosted, closed-weight alternatives in the same family, see Muse Wonder for coding work and Muse Image for image generation and editing.
This About section is AI-generated from public sources via VeniceStats + Venice inference, with no human editing. It may contain inaccuracies.
Daily usage · Muse Spark Open · Organic sellers
| Seller | Health | Rep↓ | Sales | Tokens | Routing | Input $/M | Cached $/M | Output $/M | Save |
|---|---|---|---|---|---|---|---|---|---|
| Super Seeder 0xd19f…41f3 | healthy | 89.22 | $0.26 | 86.85k | #2 | $1.0673 | $0.1281 | $3.629 | −15% |
| ▲ Apex Ant 0x73b4…e736 | healthy | 81.71 | $0.12 | 64.03k | #1 | $0.78 | $0.09 | $2.6563 | −38% |
| OpenCheap 0x0ee1…6c41 | — | 12.43 | — | — | gated | $1.046 | $0.1255 | $3.5564 | −16% |
| Fire Ant 🔥🐜 0xbe05…bc5d | — | 0.00 | $0.02 | 21.01k | gated | $1.3638 | $0.1637 | $4.6368 | — |
"Best price" and the seller table are live Antseed catalog data (advertised $/1M tokens — or $ per generated image for unit-billed image models — not settled amounts). Reputation = buyer trust score (0-100, the Antseed SDK's own formula). "Routing" = the SDK's default buyer routing (what the VPR desktop app ships with): a trust ≥ 60 gate on the effective reputation, then cheapest-first among routable sellers; live failover state (per-peer cooldowns) is buyer-side runtime and not included. Model knowledge (TLDR, provider, About) via the VeniceStats enrichment layer. Advertised catalog, not the model used in any specific purchase. "Usage on Antseed" counts only settlements whose buyers share the per-model split on-chain (metadata v2/v3, opt-in), so every usage figure is a lower bound.