Meta releases Muse Spark 1.1 model

Meta Platforms on July 9 opened public preview access to Muse Spark 1.1 for U.S. developers through its Meta Model API, expanding beyond the private preview it began after the model family’s April debut. Reuters reported the company released the upgraded model alongside developer access, while Chief Executive Mark Zuckerberg said in a post that Muse Spark 1.1 is aimed at agentic and coding tasks and offered at low cost. Meta said Muse Spark 1.1 can write and debug code, use software and external tools, process text, images and video, and carry out multi-step tasks with less human intervention. The company is also making the model available in “Thinking mode” in the Meta AI app and on the web, according to Reuters. (ca.finance.yahoo.com) ### Where does Muse Spark 1.1 actually sit in Meta’s AI lineup? Meta introduced the original Muse Spark on April 8 as the first model from Meta Superintelligence Labs and said at the time it would power the Meta AI app and website, with rollout planned across WhatsApp, Instagram, Facebook, Messenger and AI glasses. The company described that first release as purpose-built for Meta’s own products and said API access would initially be limited to select partners. (ca.finance.yahoo.com) April 8 matters because it shows the shift in Meta’s sequencing: product deployment came first, broader developer access later. Reuters said Thursday’s launch turns that internal product model into a paid external service, putting Meta into more direct competition with OpenAI and Anthropic in selling model access. (about.fb.com) ### What are developers getting, and what does it cost? U.S. developers can now use Meta Model API in public preview to test prompts, compare outputs and build prototype integrations, Reuters reported. Meta is giving new users $20 in credits before they move to pay-as-you-go pricing. (about.fb.com) The pricing Meta disclosed is $1.25 per million input tokens and $4.25 per million output tokens. CNBC reported Meta AI chief Alexandr Wang described the pricing as “very aggressive and attractive” relative to rival offerings. ### How does this connect to Zuckerberg’s September chip timeline? (ca.finance.yahoo.com) Reuters reported separately on July 9 that Meta plans to start manufacturing an in-house AI chip from September, citing an internal memo it reviewed. The report said the chip effort is part of a wider plan to increase the company’s total computing power to 14 gigawatts by 2027. (ca.finance.yahoo.com) The same Reuters report said the chip is part of Meta’s in-house accelerator roadmap, a supply and cost effort that would reduce dependence on outside chip vendors as the company expands AI infrastructure. Zuckerberg referenced both the September production target and the 14-gigawatt goal in his post about Muse Spark 1.1, tying the model launch to a broader build-out of computing capacity. (thestar.com.my) ### What has Meta said about safety and deployment risks? Meta said in a May 26 safety and preparedness report that Muse Spark had been evaluated under its Advanced AI Scaling Framework before release inside Meta AI. The company said pre-mitigation testing found elevated risk in some domains, including chemical and biological capabilities, but concluded residual risk was acceptable after safeguards were applied. (thestar.com.my) The same report said Muse Spark remained vulnerable to adaptive jailbreak and prompt-injection attacks in agentic settings and described those issues as active areas of research. Meta also said third-party testing by Apollo Research found a high rate of evaluation awareness in the model. (ai.meta.com) ### What happens next? September is the next named milestone in Meta’s AI build-out, with Reuters reporting that month as the target for chip manufacturing to begin. On the software side, Muse Spark 1.1 is now in public preview for U.S. developers through Meta Model API, while Meta has said Muse Spark features are rolling across its apps and AI glasses. (thestar.com.my) (ai.meta.com)