Muse Spark 1.2
Meta's Muse Spark 1.2 is a coding model co-trained with Muse Code, offering long-horizon training, self-improvement, and goal-conditioned planning.
Model Overview
Capabilities, design details, and architectural traits
Muse Spark 1.2 - Co-Trained Long-Horizon Coding Model
Muse Spark 1.2 is Meta's coding-focused language model, designed specifically for long-running, multi-step software engineering across large repositories rather than short autocomplete suggestions. It is the model powering Muse Code, Meta's terminal coding agent.
| Trait | Detail |
|---|---|
| Co-training with Muse Code | Trained alongside the Muse Code agent using rejection-sampled harness trajectories and recipe optimizations for goals, compaction, and subagents, plus integration of the Muse Code toolset for harness compatibility |
| Self-improvement loop | Muse Spark 1.1 generated challenging coding environments and instruction-following templates, then graded candidate solutions to produce training data for 1.2 |
| Long-horizon training | Extensively trained on whole-repository generation, large end-to-end projects, and auto-research tasks |
| Goal-conditioned planning | Uses planning to sequence work, goal conditioning to maintain direction, and context compaction to retain knowledge across extended sessions |
Co-Training and Self-Improvement
The model's training pipeline is unusual in two ways. First, Muse Spark 1.2 was co-trained with Muse Code so the model performs best when paired with that specific agent runtime. Second, its predecessor Muse Spark 1.1 was used to generate and grade training environments, creating a self-improvement loop that improved instruction-following precision.
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 | August 5, 2026 | September 22, 2026 | September 28, 2026 |
| Knowledge Cutoff | - | - | Jun 2026 |
| Context & Limits | |||
| Context Window | 1M | 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 | textimagefilevideo | textimagefile | textimagefile |
| Outputs | text | text | text |
| Benchmarks (0-100) | |||
| Intelligence Index | 39.6 | 57.6 Best Intelligence Index | 56.0 |
| Coding Index | 72.2 | - | - |
| Agentic Index | 43.2 | - | - |
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
