Meta Superintelligence Labs released Muse Spark Open on October 15, 2026: open weights for an 816B-parameter sparse mixture-of-experts model that activates about 42B parameters per token, reads text, images, video and audio, and holds a 1,048,576-token context. Meta pitches it at coding and long-horizon agent work.
The hosted model most developers already call is Muse Spark 1.3, released September 2, 2026 on Muse Code and the Meta Model API. It is a separate line from the open checkpoint, and its pricing, rate limits and API details are covered further down.
The announcement is Introducing Muse Spark Open; weights and the model card are at meta-models/Muse-Spark-Open-816B-A42B, with a collection page and developer docs.
Key Specs
| Item | Value |
|---|---|
| Parameters | 816B total / ~42B active |
| Experts | 256, top-8 routing, plus a perception encoder |
| Context | 1,048,576 tokens |
| Modalities | Text, image, video, audio in; text out |
| Checkpoints | BF16 (~1.65 TB class) and NVFP4 (~537 GB; Meta cites 700GB+ VRAM to serve) |
| License | Muse Model License, with limited-commercial and MaaS revenue-threshold terms |
Read the license before you build on it. It is not Apache or MIT.
Deployment Options
Hosted: Meta's post names Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together AI. On October 16, 2026, OpenRouter listed meta/muse-spark-open at $1.25 per million input tokens, $4.25 output and $0.15 for cached reads. Those are aggregator prices, not a Meta rate card.
Self-hosted: SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth are all called out, and Meta describes an MLX setup across two M5 Ultra Mac Studios.
Muse Spark 1.3 Pricing and API
Muse Spark 1.3 (muse-spark-1.3) is the version Meta's model docs recommend for new work. It takes text, image, video, audio and PDF input, returns text, and shares the same 1,048,576-token context as the rest of the Muse Spark family. Meta's docs flag audio understanding on 1.3 as not fully supported yet and point audio jobs to Muse Spark 1.2 or Muse Voice Transcribe.
Meta's pricing and rate limits page lists two modes, priced per million tokens:
| Mode | Model ID | Input | Cached input | Output | Rate limit |
|---|---|---|---|---|---|
| Standard | muse-spark-1.3 |
$1.25 | $0.15 | $4.25 | 3,000 RPM / 4M TPM |
| Contributor | muse-spark-1.3-contributor |
$0.10 | $0.002 | $0.20 | 150 RPM / 3M TPM |
Contributor mode is the same model at a lower price, in exchange for letting Meta use your content to train future models. It also drops the max reasoning effort, which only Standard mode supports. Vercel's AI Gateway changelog notes that both rates are unchanged from Muse Spark 1.2.
The API is reached at https://api.meta.ai/v1 and inside Muse Code on pay-as-you-go or any plan. Meta's launch post describes 1.3 as trained on more long-horizon coding tasks; in comparisons by Meta engineers it used about 20% fewer tool calls and about 25% fewer tokens than 1.2. Those are Meta's internal figures, and the post does not publish a public benchmark table alongside them.
Muse Glimmer: Meta's 30B Model for Local Agents
Muse Spark Open was not Meta's first open-weight Muse release. On August 10, 2026, Meta Superintelligence Labs published Muse Glimmer, a 30B model distilled from Muse Spark and built to run agents on a single consumer GPU or a Mac. Its knowledge cutoff is January 4, 2026. It is under Apache 2.0, a more permissive license than the Muse Model License that covers Spark Open. The announcement is Introducing Muse Glimmer and the weights are at meta-models/Muse-Glimmer-30B.
| Muse Glimmer-30B | Muse Spark Open | |
|---|---|---|
| Released | Aug 10, 2026 | Oct 15, 2026 |
| Architecture | Dense transformer, about 29.6B parameters including a 1.8B vision encoder | Sparse MoE, 816B total, about 42B active |
| Inputs | Text and images (video only as individual frames) | Text, images, video, audio |
| Context | 131,072+ tokens | 1,048,576 tokens |
| License | Apache 2.0 | Muse Model License |
According to the model card, a roughly 4-bit quantization brings the language model under 20 GB, so it fits in 24 GB or 32 GB of VRAM. Meta measures 0.2% average degradation for its K-Quant-Dynamic build (32 GB target) and 1.0% for K-Quant-17GB (24 GB target). A bundled DFlash speculative-decoding drafter raised the 17 GB build from 74.9 to 233.4 tokens per second on an RTX 5090 in Meta's tests, and from 26.6 to 50.2 on an Apple M5 Max. Reasoning strength is set in the system prompt (Reasoning strength: low, medium, high or xhigh).
Meta's card compares it with Gemma4-31B and Qwen3.6-27B in thinking mode. Glimmer leads on MCP Atlas (75.5 vs 54.2 and 62.5) and SWE-Bench Pro (51.2 vs 36.9 and 50.2). Qwen3.6-27B leads on OSWorld-Verified (75.6 vs 65.9) and Terminal-Bench 2.1 (60.7 vs 51.7). These are Meta's own runs.
On OpenRouter, meta/muse-glimmer-30b lists $0.30 per million input tokens and $1.20 per million output tokens with a 131,072-token context. Its usage fell 76% in the October 10–16 window to 9.41B tokens (rank 171). On AI Rank's Arena overall board published October 15, Muse Glimmer is rank 124 at 1424.5 from 3,897 votes.
→ Track it on AI Rank: Muse Glimmer: Arena rating, rank and 90-day history
Muse Spark 1.x and Muse Wonder Distinction
The meta/muse-spark-1.3 models on OpenRouter belong to the hosted 1.x line, which predates the open weights. Muse Wonder is a separate Meta product line; see Muse Wonder pricing and plans.
On AI Rank's OpenRouter text usage ranking for Oct 10–16, 2026, Muse Spark 1.3 Contributor ranked No. 9 with 2.93T tokens (+81%), and Standard-mode Muse Spark 1.3 ranked No. 44 with 320B (−19%). Muse Spark Open entered at No. 136 with 23.7B tokens in its first partial week, and Muse Wonder entered at No. 59 with 201B.
→ Live usage: see the OpenRouter model usage ranking on AI Rank (tokens, Oct 10–16, 2026 window).
Compare before you choose: pricing and rankings move daily — check the current model rankings on AI Rank before you commit.
Or 📱 get the AI Rank app to follow them on your phone.
Prepared with automated editorial assistance and checked against official sources and dated AI Rank data on October 16, 2026; Muse Glimmer section and usage figures updated on October 22, 2026. We did not run our own benchmarks for this article.