Thinking Machines has released Inkling 2 Small, a multimodal mixture-of-experts model with 594 billion total parameters and 17 billion activated per inference. It supports text and image inputs, a 1-million-token context window, and adjustable reasoning effort. The model is currently available for inference and fine-tuning through Tinker; following external safety testing and red-team evaluations, its weights are planned for release under the Apache 2.0 license next week. It scored 31 on the Artificial Analysis Intelligence Index.

Some investors considering participating in Anthropic's IPO believe a $1.5 trillion valuation is too high and favor a valuation of about $1 trillion to $1.1 trillion. The company had initially expected to raise up to $100 billion but now plans to lower that to about $80 billion; the valuation still depends on investors' assessment of the listing plans.

OpenAI and Ligo introduced BioBasicsBench, which uses 100 real wet-lab scientific tasks to assess AI agents' ability to interpret experimental data, identify errors, and evaluate scientific conclusions. Tasks involve gel images, microscopy videos, and sequencing reads. GPT-6 Astra scored 66.7%, compared with 43.8% for GPT-5.6 Sol; the former used about half as many output tokens per task as the latter.

An independent METR evaluation of Anthropic models found that their performance on well-defined tasks is improving at roughly twice the rate observed during 2024–2025. The evaluation identifies AI's growing role in R&D as the main driver of the acceleration, with other factors also contributing in late 2025.

Artificial Analysis's AA-AgentPerf benchmark showed that, when running DeepSeek V4.1 Flash, Vera Rubin NVL72 achieved 5x the single-GPU throughput of the GB300 configuration and 1.47x the single-agent speed. A separate test on GB300 NVL72 found that upgrading vLLM, using FlashInfer autotuning, and integrating NVIDIA Dynamo increased single-agent token speed to as much as 1.5x its original level.

Meta announced that Muse and Meta AI will use specialized models trained with Muse Wonder, adding Active Memory and Active Focus. These capabilities are designed to retain relevant information over time and keep pursuing goals when obstacles arise or conditions change, respectively. Meta's tests found that Muse was about 2.4 times less likely to forget key information and improved its handling of changing constraints by 67%.

Google donated Agent Substrate, gVisor, and OpenRL to the CNCF Sandbox and announced that Agent Substrate is now available on Google Kubernetes Engine. The infrastructure can restore idle sessions within 500 milliseconds, increase container density by 10x, and support more than 500 suspend/resume operations per second, providing infrastructure for agent execution and isolation.

Zyphra open-sourced its ZYFA kernel library, which implements FP8 and BF16 causal forward attention on AMD Instinct MI355X. Tests showed real-model activation throughput at 1.16 to 1.76 times that of AITER; at a head dimension of 128, FP8 error was reduced by 62% to 78%. The two implementations were tested at different operating frequencies. The code is licensed under Apache 2.0.

Anthropic outlined Claude's memory-engineering approach, including what information to select for memory, when to consolidate it, and how to organize it. The company is also exploring the use of idle compute to consolidate and correct existing memories. In a test using 350,000 documents, the score rose from 13% to 48% after 128 consolidation cycles, and reached 61% when combined with search and a Wiki. These are test results, and performance depends on the task and how the memory is organized.

Google Antigravity has opened access to remote agents, which can run user-submitted tasks in isolated cloud VMs. Users can manage permissions from a mobile device and later receive test results and code merge requests. The feature is available with all paid individual subscription plans. Antigravity also supports connecting GitHub repositories and Google Drive folders within the same project, enabling users to work with code and documents across web and mobile apps.