Muse Spark Open
Meta Muse Spark Open is an 816B sparse MoE open-weight model with 42B active parameters, 1M-token context, multimodal input, and long-horizon agentic coding.
Model Overview
Capabilities, design details, and architectural traits
Muse Spark Open - Meta's open-weight agentic MoE model
Muse Spark Open is the open-weights release of Muse Spark, built by Meta Superintelligence Labs and positioned by Meta as the most capable US open-weight model to date. It is a sparse mixture-of-experts model trained specifically for long-horizon agentic work such as coding and customer service agents, and it shares the same base model as Muse Spark 1.3.
| Trait | Detail |
|---|---|
| Sparse MoE scale | 816B total parameters with roughly 42B active per token, released under the Muse Model License |
| 1M-token context | Holds a million tokens, enough to read an entire codebase with dependencies, legacy code and thousands of files without splitting the task |
| Native multimodality | Takes text, image, video and audio input through a NaViT visual encoder and a dedicated audio encoder feeding a shared language decoder |
| Long-horizon agentic training | Trained to read a large codebase, plan a change, call tools, check results and self-correct when a step fails, not just single-shot tool calls |
| Subagent and computer-use training | Trained to spawn and coordinate subagents, browse the web including multimodal browsing, and use desktop, mobile and browser in a unified computer-use capability |
| Intent commentary | Generates separate commentary stating intent before acting, so users always know what the model is doing and why |
| Always-on reasoning with adjustable effort | Reasons before every answer, with reasoning effort levels from low to xhigh |
Built for real developer workflows
The model is trained on real developer workflows such as multi-file refactors, debugging sessions and long-running tasks, aiming at higher first-attempt accuracy and reliable tool calling. It received additional safety post-training to make it more robust as a standalone model without system guardrails.
Open post-training path
TorchTitan supports supervised fine-tuning, LoRA and reinforcement learning out of the box.
Benchmark Performance
Independent evaluations · Artificial Analysis
Accuracy & Capability Details
Compare Models Side-by-Side
Evaluate specifications, pricing, and independent benchmark indices
| Model Details | |||
|---|---|---|---|
| General Info | |||
| Provider | Meta | Anthropic | Anthropic |
| Release Date | October 15, 2026 | September 22, 2026 | September 28, 2026 |
| Knowledge Cutoff | - | - | Jun 2026 |
| Context & Limits | |||
| Context Window | 1.0M Best Context Window | 1M | 1M |
| Pricing (per 1M tokens) | |||
| Input Pricing | $1.25 Best Input Pricing | $4 | $2 |
| Output Pricing | $4.25 Best Output Pricing | $20 | $10 |
| Modalities | |||
| Inputs | textimagefileaudiovideo | textimagefile | textimagefile |
| Outputs | text | text | text |
| Benchmarks (0-100) | |||
| Intelligence Index | 38.9 | 57.6 Best Intelligence Index | 56.0 |
| Coding Index | - | - | - |
| Agentic Index | - | - | - |
Humanity's Last Exam
Extremely difficult logical reasoning and knowledge.
Long Context Reasoning
Logical reasoning over long context windows.
SciCode Benchmark
Scientific coding and mathematical modeling.
Independent evaluation data provided by Artificial Analysis. To view the latest benchmarks and full details, visit their official site.
Explore more from Meta
Other models by Meta
Top AI Models
Leading alternatives by intelligence score
Claude Opus 5.5
GPT-6 Astra
Gemini 4 Argon
