Muse Spark 1.2: Technical Specifications, Pricing, and Context Length

Meta’s Muse Spark 1.2 is a specialized foundation model engineered specifically for autonomous software engineering, repository-scale refactoring, and multi-step agent trajectories. Launched on August 5, 2026 alongside the Muse Code terminal agent, Muse Spark 1.2 was co-trained directly with its agent runtime, incorporating context compaction, goal conditioning, and subagent worktree orchestration into the underlying weights.

On standardized benchmarks, Meta reports Muse Spark 1.2 achieving 82.9% on Terminal-Bench 2.1, surpassing Claude Sonnet 5 (80.4%) by +3.1% relative and scoring an industry-leading 90.3% on MCP-Atlas tool execution. Explore where Muse Spark ranks in our AI Model Leaderboard and Full LLM Model Directory.

Executive Summary & Core Specification Matrix

Architectural Advancements: Model-Harness Co-Training

The critical technical differentiator in Muse Spark 1.2 is co-training with the Muse Code runtime:

1. Context Compaction: Rather than carrying an ever-expanding conversational history that degrades attention fidelity, Muse Spark 1.2 was trained to compress past tool executions, bash outputs, and compiler traces into compact state representations.
2. Subagent Worktree Coordination: The model understands Git worktree isolation, allowing it to delegate sub-tasks to parallel agent workers without generating conflicting file modifications.
3. Continuous Execution Persistence: In an automated optimization experiment, Muse Spark 1.2 executed over 1,000 continuous tool calls across a 24-hour window, writing, profiling, and optimizing custom Triton GPU kernels on NVIDIA Hopper hardware without state collapse.

graph TD
    A["Developer Goal"] --> B["Muse Code Runtime"]
    B --> C["Muse Spark 1.2 Engine"]
    C --> D["Context Compaction & Planning"]
    D --> E["Tool Calls & Subagent Execution"]
    E --> F["Repository Diff & Test Suite"]
    F --> G["Final Code Verification"]

Benchmark Comparisons: Terminal-Bench 2.1

Meta evaluated Muse Spark 1.2 against top frontier models on Terminal-Bench 2.1:

Methodological Note: Scores represent vendor-reported evaluations. Terminal-Bench evaluates live command-line tool execution, flag usage, and error recovery in real bash environments. For standard benchmarks, see our SWE-bench Verified Guide.

Token Economics: Standard vs Contributor Tier

Meta provides two commercial access tiers through the Meta Developer API:

• Standard Developer Tier: $1.25/M input and $4.25/M output, offering standard zero-retention enterprise privacy terms.
• Discounted Contributor Tier: $0.10/M input and $0.20/M output (a 92% input discount and 95% output discount). In exchange, Meta receives telemetry and interaction data to train future models.

For independent developers, open-source projects, and exploratory scripting, the Contributor tier provides exceptional value. For proprietary corporate codebases, standard rates or self-hosted models like Muse Glimmer remain required.

Final Verdict

Meta Muse Spark 1.2 represents the vanguard of harness-co-trained coding models.

By pairing an 82.9% Terminal-Bench score with 1M context length, context compaction, and $1.25 / $4.25 pricing, Meta delivers an engine purpose-built for the long-horizon autonomous workflows that define modern software engineering.

Read our complete How to Use Meta Muse Code Guide and track real-time benchmark scores on our AI Model Leaderboard.