Muse Spark

Muse Spark is Meta's proprietary family of frontier large language models from Meta Superintelligence Labs. The current version, Muse Spark 1.2, launched on August 5, 2026 as a coding-focused update to Muse Spark 1.1. It is built for software engineering, multimodal reasoning, computer use, tool-using AI agents, and projects that continue across many steps rather than ending after one answer. Muse Spark 1.2 also powers Muse Code, Meta's terminal coding agent.

The simplest distinction is that Muse Spark is Meta's larger, cloud-hosted model. Developers access it through the Meta Model API or through Muse Code; Meta does not publish downloadable Muse Spark weights. Muse Glimmer, by contrast, is a smaller 30B open-weight model distilled from Spark for local deployment.

What Muse Spark can do

Muse Spark 1.2 concentrates on code generation, complex debugging, large-codebase understanding, repository-wide changes, and end-to-end developer workflows. Meta says it increased the amount of coding compute and the diversity of training environments used for 1.2 while preserving the broader agent capabilities of Muse Spark 1.1. Those capabilities include planning, calling tools, operating software, interpreting images and documents, coordinating subagents, and maintaining direction over long tasks.

Muse Spark 1.2 was co-trained with the Muse Code harness. Training included trajectories from the agent's tools plus optimizations for planning, goals, context compaction, and subagent delegation. That pairing matters when reading benchmark results: a score from Muse Spark inside Muse Code measures the model and Meta's agent runtime together, not the model in isolation.

Meta demonstrated the long-horizon design with GPU-kernel optimization runs lasting more than 1,000 tool calls and as long as 24 hours. A demonstration is not the same as a general reliability guarantee, but it shows the type of sustained engineering work the model is intended to handle.

Context window and multimodal input

Muse Spark 1.2 retains the one-million-token context window introduced with Muse Spark 1.1. A context window that large can hold extensive repositories, long documents, tool histories, images, video-derived information, and earlier decisions from an agent run. Meta also trains the model to compact its working history and retrieve important earlier actions instead of treating the entire million-token window as equally useful at every step.

The model family is multimodal. Muse Spark 1.1 accepts and reasons over text, images, video, PDFs, and computer interfaces, and 1.2 is presented as retaining those general-agent strengths while improving coding. A large context limit does not guarantee perfect recall, and multimodal support does not mean every input type performs equally well. Production systems still need file selection, retrieval, permissions, and checks around tool actions.

Meta has not publicly disclosed Muse Spark's parameter count, detailed architecture, training-data mixture, or downloadable checkpoint. The model therefore cannot be independently inspected or self-hosted in the way an open-weight model can.

Benchmark profile

Meta's official launch comparison placed Muse Spark 1.2 against frontier coding models, each running in its vendor's preferred agent harness.

Official launch evaluation	Muse Spark 1.2	Claude Opus 5	GPT-5.6 Terra	Read
Terminal-Bench 2.1	82.9%	86.7%	81.8%	Spark ranks second in Meta's chart.
DeepSWE 1.1	59.3%	65.0%	64.8%	Spark ranks third among these models.
Meta Internal Coding Bench	70.6%	79.4%	65.4%	Opus leads even on Meta's internal test.

The table is transcribed from Meta's official launch charts and methodology material. Spark ran in Muse Code, Claude in Claude Code, and GPT in Codex, so the results measure each model-plus-harness combination rather than isolating the underlying models. Meta also co-trained Spark with Muse Code, making the harness part of the product advantage being evaluated.

Independent testing gives a broader, more reproducible view.

Artificial Analysis evaluation	Muse Spark 1.2	Comparison or context
Intelligence Index at launch	54	Spark 1.1 scored 51.
GDPval-AA v2	1,631 Elo	Fifth among models tested at launch; ahead of Claude Opus 4.8 at 1,588.
Terminal-Bench v2.1	80%	Up from 78% for Spark 1.1.
tau3-Banking	27%	Up from 25% for Spark 1.1.
Intelligence Index task cost	USD 0.40	Based on standard USD 1.25 / USD 4.25 token pricing.
AA-Omniscience hallucination rate	28%	Down from 38% for Spark 1.1; lower is better.

Artificial Analysis later listed Spark 1.2 at 57 after an index update, a reminder that leaderboard scores can change when graders or index versions change. The independent results show a material improvement in agentic knowledge work without placing Spark ahead of every frontier model.

API access, pricing, and data use

Muse Spark 1.2 is available through the Meta Model API and Muse Code. Standard API pricing at launch is USD 1.25 per million input tokens, USD 4.25 per million output tokens, and USD 0.15 per million cached input tokens. Meta states that standard-tier prompts and completions are not used to train its models.

Meta also offers a lower-priced Contributor tier: USD 0.10 per million input tokens, USD 0.20 per million output tokens, and USD 0.002 per million cached input tokens. In exchange, users give Meta permission to use their prompts and completions for model training. That makes the contributor tier inappropriate for proprietary code, confidential documents, personal data, or regulated workloads unless the organization has explicitly accepted that use. The standard and contributor tiers are the same model but have different data terms, prices, and rate limits.

Muse Spark is not a free local model. It requires an API account and paid usage, even when accessed through the discounted contributor tier. Prices and availability can change, so production buyers should verify the current Meta Model API documentation.

When to use it

Muse Spark 1.2 is a candidate for repository-scale coding, complex debugging, software migrations, multimodal computer-use agents, research workflows, and long-running tasks that benefit from a very large context window. Muse Code is the most integrated way to use its coding behavior because the model and harness were trained together.

When not to use it

Muse Spark is not the right choice when downloadable weights, offline operation, full architecture transparency, or local data control are requirements. The contributor API should not be used for sensitive material when training-data reuse is unacceptable. Teams that mainly need short classification or simple chat may also pay for agent-oriented capabilities they do not use.

Muse Spark vs. Muse Glimmer

Muse Spark is the more capable proprietary teacher model, accessed through Meta's cloud. Muse Glimmer is the smaller 30B student model, released under Apache 2.0 and designed to run on a high-end Mac or a single consumer GPU. Spark offers the stronger frontier-level service and a one-million-token context; Glimmer offers downloadable weights, local privacy, customization, and a 131K-plus context window.

Bottom line

Muse Spark 1.2 is Meta's cloud model for serious coding and agent work, not a continuation of Llama's downloadable-weight strategy. Its strongest practical features are the million-token context window, close integration with Muse Code, and relatively low standard API price. The tradeoffs are equally important: its architecture is undisclosed, its weights cannot be downloaded, and its cheapest API tier exchanges a steep discount for permission to train on user data.