Meta released the weights of its Muse Spark model, proprietary since April, as Muse Spark Open, which it calls the most capable open-weight model yet from a U.S. developer, per SiliconANGLE. Meta said the model passed the release bar set by its Meta Superintelligence Scaling Framework and that it will publish a full evaluation report later. The new license rules out military and critical infrastructure uses and requires companies that sell hosted access and earn more than $20 million in annual revenue to sign a separate deal. In Meta's own benchmark table, Chinese open models led on coding, with Moonshot's Kimi K3 scoring 85 to Muse Spark Open's 76.4. Meta worked with Apple to run the model on two Mac Studio desktops. "Meta has no control or visibility" into such local use, Meta AI chief Alexandr Wang told Axios.
Read at SiliconANGLE ↗ • Read at Axios ↗
Perplexity has open-sourced SPACE, its sandboxSandboxIsolated computing environment where an agent can run code and use tools without touching the machine underneath for isolating AI agentsAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise., per SiliconANGLE. CoreWeave, Akamai and Crusoe, which run cloud platforms for AI, already use it. The release follows several incidents in which agents escaped sandboxes, including OpenAI test agents that broke out and hacked into Hugging Face. SPACE runs each agent in its own small virtual machine and holds network access and credentials outside it, so agents can reach authorized services without seeing the credentials. Perplexity engineer Zibi Braniecki told Semafor that nine frontier models instructed to break out of SPACE failed in all 118 attempts, though some got around network restrictions that Perplexity has since fixed. He said the company has focused on hardening guardrails instead of training models to behave.
Read at SiliconANGLE ↗ • Read at Semafor ↗
OpenAI launched a Rosalind Academic Plan that gives eligible academic labs up to 30 free business seats for as long as 12 months, the company said. Access to GPT-Rosalind, its specialized life sciences model, still requires approval through OpenAI's trusted access process, which the company says lets vetted researchers pursue sensitive scientific questions within safeguards. Listed research areas include drug and protein design and antibody and vaccine work against emerging viral variants. Research data shared through the plan is not used to train OpenAI's models by default.
Read at OpenAI ↗
A preprint posted to arXiv by Morio Hamada, Magda Dubois and seven co-authors finds that transcripts of how models solved benchmark tasks can still gauge capability after scores hit their ceiling. The authors say benchmark saturation otherwise weakens AI safety cases (structured arguments that a model is safe to deploy) and frontier AI policies. Among the methods tested, pairwise judging, which compares two models' transcripts side by side, cut prediction error 5.3 times compared with using saturated scores alone. Most of the signal sits in the models' reasoning and tool calls. The methods did not always carry over to new conditions, underestimating the first reasoning models released in early 2025.
Read at arXiv ↗ • Read at alphaXiv ↗