Muse
Meta's current line, built for long multi-agent runs.
Muse is what Meta Superintelligence Labs ships now, and the family to look at rather than Llama if you want Meta's recent work. The Spark models are multimodal reasoning models for long-running agentic and multi-agent workflows — they read text, images, video, audio and files with a million-token context, and are built to hold onto what they learned early in a long task. Each Spark release also has a cheaper "Contributor" tier meant for experimenting rather than production. Glimmer 30B is the odd one out and the interesting one: open weights, distilled from Spark, small enough to run autonomous agents on ordinary hardware.
At a glance — Muse Spark 1.1
- Cost
- $1.25 in · $4.25 out per 1M tokensMid-range
- Memory
- 1M tokens — about 1,550 pages
- Reads
- Text · Images · Files · Audio · Video
- Can do
- Use tools (agents)
- Structured output
- Thinking mode
- Good at
- Coding: better than 79% · Agentic work: better than 60% · Overall: better than 61%
- Reliability
- 100.00% uptime, last 24h
- Released
- Jul 16, 2026
Advanced details
- API model ID
meta/muse-spark-1.1- Max output
- 944K tokens
- Cached input
- $0.15 per 1M
- Tokenizer
- Other
- Coding index
- 71.3
- Agentic work index
- 25.8
- Overall index
- 33.7
Providers
The same model, hosted by different companies. OpenRouter picks one per request and falls back to the next if it fails.
| Provider | Input /1M | Output /1M | Cached /1M | Max output | Uptime 24h |
|---|---|---|---|---|---|
| Meta | $1.25 | $4.25 | $0.15 | 944K | 100.00% |
Supported parameters
- include_reasoning
- max_tokens
- reasoning
- reasoning_effort
- repetition_penalty
- response_format
- structured_outputs
- temperature
- tool_choice
- tools
- top_k
- top_p
Live from OpenRouter, refreshed hourly. Scores from Artificial Analysis.
Key info
- Pricing
- Paid
- Category
- Models
- Best for
- Expert
- Made by
- Meta Superintelligence Labs
- Weights
- Closed, except Glimmer 30B
- Good for
- Long multi-agent runs over mixed media