# Muse Spark 1.2 ![Organization](https://img.shields.io/badge/Org-Meta-1877f2) ![Pricing](https://img.shields.io/badge/Price-Meta%20API-2da44e) ![License](https://img.shields.io/badge/License-Proprietary-d29922) ![Modality](https://img.shields.io/badge/Modality-Multimodal-8957e5) ![Released](https://img.shields.io/badge/Released-Aug%202026-555) > **Quick answer:** Muse Spark 1.2 is Meta's August 2026 coding-focused update to Muse Spark 1.1, released alongside Muse Code, a new terminal coding agent. It posts 90.3% on MCP Atlas — the highest score of any model Meta compared — plus 82.9% on Terminal-Bench 2.1, 59.3% on DeepSWE 1.1, and a 1,631 Elo rating on GDPVal-AA v2. Proprietary, served via the Meta Model API. ![Muse Code and Muse Spark 1.2 launch artwork](https://research.meta.ai/media/muse-code-spark/hero/dark/v2?w=960&cache=1) ## At a Glance **Where Muse Spark 1.2 leads** - 90.3% MCP Atlas — the highest score of any model in Meta's comparison set, ahead of Claude Opus 5 (xhigh, 85.8%), Gemini 3.5 Flash (83.6%), Claude Fable 5 (83.3%), Kimi K3 (82.3%), and GPT-5.6 Sol (81.8%) - Clear generational jump over Muse Spark 1.1 on every benchmark Meta reports: Terminal-Bench 2.1 (82.9% vs 76.2%), DeepSWE 1.1 (59.3% vs 53.0%), GDPVal-AA v2 (1,631 vs 1,371 Elo), MCP Atlas (90.3% vs 88.1%) - Extensively trained on long-horizon coding: whole-repository generation, large end-to-end projects, and auto-research workflows **Where it lags** - Trails Claude Opus 5 (max, Claude Code) on Terminal-Bench 2.1 (82.9% vs 86.7%) and DeepSWE 1.1 (59.3% vs 65.0%) - Also trails GPT-5.6 Terra (max, Codex) on DeepSWE 1.1 (59.3% vs 64.8%), though it edges GPT-5.6 Terra on Terminal-Bench 2.1 (82.9% vs 81.8%) - Proprietary — no open weights, Meta API access required **Best for:** Terminal-based agentic coding via Muse Code; MCP tool-use-heavy workflows; long-horizon, whole-repository coding tasks. ## What Muse Spark 1.2 Is Muse Spark 1.2 is Meta's coding-focused update to Muse Spark 1.1, released August 5, 2026 alongside Muse Code (beta) — a new terminal coding agent that pairs a simple agent loop with persistent async background agents. These background agents stay active for the whole session rather than being spawned per task, carrying out next steps and choosing when to report back to the main agent, which reduces redundant information-gathering and steering overhead on long, difficult tasks. Muse Code also uses an append-only local event log for every model call, tool run, approval, and edit, making sessions replay-exact and restart-safe after a crash. Meta co-trained Muse Spark 1.2 together with Muse Code specifically so the two perform best as a pair — training included rejection-sampled harness trajectories and recipe optimizations for goal-setting, context compaction, and subagent coordination, plus integration of the full Muse Code toolset for maximum harness compatibility. The model was also extensively trained on long-horizon coding tasks (whole-repository generation, large end-to-end projects, auto-research), using planning to sequence work, goal conditioning to hold direction, and context compaction to sustain progress across long sessions. A self-improvement loop — using Muse Spark 1.1 to generate challenging coding environments and instruction-following templates, then grading candidate solutions against those requirements — produced additional training data that improved Muse Spark 1.2's ability to follow complex instructions precisely. Meta reports Muse Spark 1.2 significantly scaled up coding-task training compute and training-environment diversity relative to 1.1, while maintaining strength in general agentic tasks. In a case study, the model iteratively optimized GPU kernels (KDA and MLA, on NVIDIA Hopper GPUs) over 1,000+ tool calls and up to 24 hours, achieving substantial speedups over baseline Triton implementations without being allowed to import third-party kernel libraries. ## Specifications | Field | Value | |---|---| | Organization | Meta | | Parameters | Undisclosed | | Context window | Undisclosed | | Architecture | Proprietary, undisclosed | | License | Proprietary | | Release date | August 5, 2026 | | Modality | Text and vision (multimodal) | ## Pricing Muse Spark 1.2 is available today through the Meta Model API with expanded global access, and via Muse Code (beta), Meta's terminal coding agent (`curl -fsSL https://dev.meta.ai/install.sh | bash`). Refer to Meta's developer site for current API rates. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [MCP Atlas](/benchmarks/scale-ai/mcp-atlas) | 90.3% | [Meta release blog](https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2) | 2026-08 | | [Terminal-Bench 2.1](/benchmarks/tbench/terminalbench-21) | 82.9% | Meta release blog (via Muse Code, xhigh effort) | 2026-08 | | [GDPval-AA v2](/benchmarks/artificial-analysis/gdpval-aa-v2) | 1,631 Elo | Meta release blog (results via Artificial Analysis) | 2026-08 | | [DeepSWE](/benchmarks/datacurve/deepswe) | 59.3% | Meta release blog (via Muse Code, xhigh effort) | 2026-08 | *Scores above are reported by Meta and shown for context. They are not Benchgen measurements. Terminal-Bench 2.1 and DeepSWE 1.1 were run with Muse Code as the agent harness (xhigh reasoning effort) in Meta's internal agent evaluation framework (isolated Daytona cloud sandboxes); MCP Atlas results come from Scale AI's own harness; GDPVal-AA v2 results come from Artificial Analysis's Stirrup harness. See the [full methodology report](https://research.meta.ai/static/muse-spark-1-2-methodology) for details. Meta also reports a proprietary "Meta Internal Coding Bench" score (70.6%, using an internal agentic harness over 440 tasks sourced from Meta's own codebase) — not included above since it has no public dataset or independently verifiable methodology.* ## Muse Spark 1.2 vs Alternatives | Model | MCP Atlas | Terminal-Bench 2.1 | DeepSWE 1.1 | GDPVal-AA v2 (Elo) | |---|---|---|---|---| | **Muse Spark 1.2** | 90.3% | 82.9% | 59.3% | 1,631 | | [Claude Opus 5](/models/anthropic/claude-opus-5) | 85.8% | 86.7% | 65.0% | 1,852 | | [GPT-5.6 Terra](/models/openai/gpt-5-6-terra) | 81.8%* | 81.8% | 64.8% | 1,577 | | [Grok 4.5](/models/xai/grok-4-5) | — | 81.6% | 56.6% | 1,526 | | [Kimi K3](/models/moonshot-ai/kimi-k3) | 82.3% | — | — | — | *GPT-5.6 Terra's MCP Atlas figure is GPT-5.6 Sol's reported score from the same chart (Meta's MCP Atlas comparison set uses "GPT 5.6 Sol" rather than "GPT 5.6 Terra" — the two entries may reflect different GPT-5.6 variants; shown here for reference only. Muse Spark 1.2 posts the single highest MCP Atlas score Meta reports (90.3%), ahead of every other model in its comparison set including Claude Opus 5 run at max effort. But on raw coding-agent benchmarks — Terminal-Bench 2.1, DeepSWE 1.1, and GDPVal-AA v2 — Claude Opus 5 (run at max effort with Claude Code) remains the strongest performer Meta compared against, with Muse Spark 1.2 in second place ahead of GPT-5.6 Terra, Grok 4.5, and its own predecessor Muse Spark 1.1 on every metric. ## How Muse Spark 1.2 Performs on Real Agent Tasks Muse Spark 1.2's MCP Atlas result stands out because it's an outright win against every other frontier model Meta tested, including models run at their own maximum reasoning settings — a strong signal for tool-use-heavy agent workloads across the 36 real MCP servers and 220 tools MCP Atlas covers. Its Terminal-Bench 2.1 and DeepSWE 1.1 results, run specifically through Muse Code (its co-trained, purpose-built harness), show the model is a clear step up from Muse Spark 1.1 on both long-horizon terminal tasks and real-world software-engineering pull-request-style work, even though it trails Claude Opus 5 on both. Builders should note that Meta's own evaluation setup — including agent tools and system prompts — may not be specifically tuned for third-party proprietary models like Opus 5 or GPT-5.6, so those comparison figures may understate their best achievable performance; Muse Spark 1.2's own numbers, run in its native Muse Code harness, are the more reliable read on the model's actual capability ceiling. ## Related Models & Benchmarks - Compare: [Claude Opus 5](/models/anthropic/claude-opus-5) · [GPT-5.6 Terra](/models/openai/gpt-5-6-terra) · [Grok 4.5](/models/xai/grok-4-5) · [Kimi K3](/models/moonshot-ai/kimi-k3) · [Muse Spark](/models/meta/muse-spark) · [Muse Glimmer](/models/meta/muse-glimmer) - Benchmarks: [MCP Atlas](/benchmarks/scale-ai/mcp-atlas) · [Terminal-Bench 2.1](/benchmarks/tbench/terminalbench-21) · [DeepSWE](/benchmarks/datacurve/deepswe) · [GDPval-AA v2](/benchmarks/artificial-analysis/gdpval-aa-v2) ## Frequently Asked Questions

What is Muse Spark 1.2?

Muse Spark 1.2 is Meta's August 2026 coding-focused model update, co-trained with Muse Code — Meta's new terminal coding agent. It scores 90.3% on MCP Atlas (the highest of any model Meta compared), 82.9% on Terminal-Bench 2.1, and 59.3% on DeepSWE 1.1.

What is Muse Code?

Muse Code (beta) is Meta's terminal coding agent, powered by Muse Spark 1.2. It uses a simple agent loop plus persistent async background agents, a replay-exact/restart-safe local event log, and bundled skills like `/plan`, `/grill`, and `/goal`. Install via `curl -fsSL https://dev.meta.ai/install.sh | bash`.

How does Muse Spark 1.2 compare to Muse Spark 1.1?

Muse Spark 1.2 improves on Muse Spark 1.1 across every benchmark Meta reports: Terminal-Bench 2.1 (82.9% vs 76.2%), DeepSWE 1.1 (59.3% vs 53.0%), GDPVal-AA v2 (1,631 vs 1,371 Elo), and MCP Atlas (90.3% vs 88.1%).

How does Muse Spark 1.2 compare to Claude Opus 5?

Muse Spark 1.2 leads Claude Opus 5 on MCP Atlas (90.3% vs 85.8%), but Claude Opus 5 (run at max effort with Claude Code) leads on Terminal-Bench 2.1 (86.7% vs 82.9%), DeepSWE 1.1 (65.0% vs 59.3%), and GDPVal-AA v2 (1,852 vs 1,631 Elo).

Is Muse Spark 1.2 open weight?

No — Muse Spark 1.2 is proprietary and served through the Meta Model API. Meta's separate open-weight local model is Muse Glimmer.

--- *Specs from Meta's Muse Code and Muse Spark 1.2 release blog and methodology report (August 5, 2026). Last updated 2026-08-12.*

Muse Spark 1.2 — Evaluations

Evaluations