Meta Opens Muse Spark 1.1 API With 1 Million-Token Agent Context
SiliconANGLE reported that Meta launched Muse Spark 1.1 in Meta AI and a public-preview Meta Model API, with a 1 million-token context window and company-reported coding benchmark gains. API pricing, named enterprise customers and independent performance validation remain outside the public report.
Muse Spark 1.1 is now available through the Meta AI chatbot and a public-preview Meta Model API, giving developers access to a new flagship Meta model for multi-agent automation workflows.
SiliconANGLE described Muse Spark 1.1 as a model built to keep more agent work inside context during long tasks.
The public-preview API allows developers to embed the large language model in custom software.
Meta Model API Opens Muse Spark 1.1 To Developers
The new model is built for workflows where one main agent plans a task and subagents carry out parts of it.
Meta stated that Muse Spark 1.1 can respond to mid-task developments that require the plan to change.
A context compaction mechanism handles data generated while agents work through multiple steps.
Meta indicated that the mechanism preserves important details and lets the model retrieve information from earlier work when it needs to move data between sub-tasks.
The context window for Muse Spark 1.1 is listed at 1 million tokens.
Context overflow was framed as a quality risk when agents generate more data than the underlying model can retain.
Coding Benchmarks Remain Meta-Reported
Engineers at Meta tested Muse Spark 1.1 by asking it to generate a chat app from prompts.
The model took screenshots of the interface, identified technical issues, found the code snippets behind them and fixed the problems.
Meta reported that Muse Spark 1.1 scored 72.2 on Vibe Code Bench v1.1, an AI programming benchmark.
In the same comparison, the model finished more than 50 points ahead of the company's previous flagship large language model, while a second test, SWE-Atlas Codebase QnA, showed a nearly 18% higher score.
Those benchmark claims remain company-provided.
Independent benchmark validation, customer deployment results and third-party testing for the new model remain outside the report.
Reuters Report Names Meta's 14-Gigawatt Capacity Plan
The report also cited a Reuters account that Meta plans to raise its data centre capacity to 14 gigawatts next year.
Reuters noted that the plan revolves around Iris, an internally developed AI chip set for mass production in September.
That account added that Iris is likely the MTIA400 chip that Meta previewed in March.
SiliconANGLE wrote that the chip includes 51% more high-bandwidth memory than Meta's previous-generation silicon and supports enhanced MX8 and MX4 data formats.
Meta stated that MTIA400 is 400% faster than its predecessor.
The remaining public-record gaps are pricing for the Meta Model API, named enterprise customers, independent performance validation and a production availability date beyond the public-preview API.
