# Meta Releases Muse Spark 1.2 Coding Model Alongside Muse Code Beta > The August 5, 2026 launch introduces a one million token context window and persistent agents designed for complex multi-hour coding projects across large repositories. *Published 2026-08-06 · By Marcus Vance* Muse Spark 1.2 is Meta's coding-optimized model co-trained with the Muse Code beta terminal agent to enable reliable performance on extended software engineering workflows. Meta announced the release of Muse Spark 1.2 and the accompanying Muse Code beta on August 5, 2026. The update represents an effort to advance capabilities in agentic coding by scaling training compute specifically for coding tasks. This approach aims to improve performance in code generation, complex debugging, and full developer workflows involving large repositories. ## What background context surrounds the Muse Spark series? Meta Superintelligence Labs has been actively working on advancing AI capabilities in various domains. A recent emphasis has been placed on agentic systems. Previous iterations of the Muse Spark model provided foundational coding support. The 1.2 version addresses limitations in handling very long contexts and maintaining consistency over prolonged sessions. The decision to release both the model and a dedicated agent reflects an understanding that effective coding AI requires tight integration between the language model and the execution environment. The timing aligns with growing industry interest in AI tools that can autonomously manage substantial portions of software development. ## What details define the new features in Muse Spark 1.2? Key updates in Muse Spark 1.2 include an expanded context window of one million tokens. This enables the model to process and reference extensive amounts of code and documentation within a single session. The model also shows improvements in first-attempt accuracy on coding problems. Tool calling has been made more reliable. These enhancements stem from the co-training process that incorporated data from actual agent trajectories. The model is optimized for whole-repository generation and handling large projects. ## How does the co-training process enhance performance? Co-training Muse Spark 1.2 with Muse Code involved several specific techniques. Rejection sampled harness trajectories were used to select only the most successful task executions for training data. Recipe optimizations adjusted the composition of the training data to include a wider variety of coding scenarios. - Utilization of rejection sampled harness trajectories from successful agent runs to filter high-quality examples. - Application of recipe optimizations to refine training data composition and focus on relevant tasks. - Direct integration of the Muse Code toolset to improve compatibility and performance in paired use. ## What capabilities does the Muse Code beta agent provide? Muse Code functions as a terminal-based agent that supports persistent asynchronous background processes. These agents continue to run and monitor tasks even when the main session is paused. An append-only local event log records all actions in a way that permits exact replay or safe restarts. The agent is optimized for software engineering tasks that involve navigating large code repositories. Its design supports sessions that can extend up to 24 hours in duration as demonstrated in internal testing scenarios. ## What are the pricing details for API access? Access to Muse Spark 1.2 through the Meta Model API follows a tiered pricing structure based on token usage. This structure is intended to make the model accessible for both small-scale experiments and large-scale production deployments. Muse Spark 1.2 pricing on the Meta Model API as of the August 2026 releaseToken CategoryPrice per 1 Million TokensCached Input$0.15Standard Input$1.25Output$4.25 ## What implications does this have for the market and stakeholders? The introduction of these tools with expanded global access could influence how software development teams incorporate AI into their processes. Companies may explore using the persistent agents for tasks that previously required significant human oversight. This potentially leads to productivity gains in enterprise environments. Meta's move challenges other providers by offering a tightly integrated model-agent pair specifically tuned for coding. This could accelerate innovation in the field as competitors respond with their own advancements in agentic capabilities. ## What expert reactions have been shared regarding the release? > Muse Spark 1.2 has improved coding capabilities compared to its predecessor. We significantly scaled up training compute on coding tasks while expanding training environment diversity, delivering improvements in code generation, complex debugging, and end-to-end developer workflows. The model was co-trained with Muse Code to optimize performance when paired together, and focused on tasks like whole-repo generation, large projects, and auto-research.AI at Meta, Official account The official statement from the AI at Meta account underscores the improvements achieved through increased compute scaling and environmental diversity in training. It also points to future plans for even larger models. ## What can be anticipated in future updates? Meta has signaled that additional models with greater capabilities are forthcoming. These future releases are expected to build upon the co-training methodology established with Muse Spark 1.2 and Muse Code. Developers should monitor the Meta Model API for updates on availability and new features. The focus on long-horizon tasks suggests that subsequent versions will aim to push the boundaries of what autonomous agents can achieve in software engineering. This includes more sophisticated research and optimization loops. ## Sources 1. [We're excited to release Muse Code (beta), a terminal coding agent powered by Muse Spark 1.2, our newest model. This marks our next step toward the frontier, with larger and much more capable models on the way.](https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2) 2. [Muse Spark 1.2 is optimized for real coding workflows, with higher first-attempt accuracy and more reliable tool calling. With 1M context, long-running tasks run start to finish in one session.](https://developer.meta.com/ai/resources/blog/build-with-muse-code/) 3. [The official statement on improvements in Muse Spark 1.2.](https://x.com/AIatMeta/status/2085084713203487041) 4. [Muse Spark 1.2 is optimized for real coding workflows, with higher first-attempt accuracy and more reliable tool calling. With 1M context, long-running tasks run start to finish in one session.](https://developer.meta.com/ai/models/muse-spark/) --- Source: https://aiintelreport.com/frontier-models/meta-releases-muse-spark-1-2-muse-code-beta Index: https://aiintelreport.com/llms.txt · Full text: https://aiintelreport.com/llms-full.txt