Thinking Machines has released Inkling 2 Small, with 594B total parameters and 17B active parameters, and plans to release its weights under Apache 2.0 next week after completing safety testing. Using internal Anthropic data, METR estimates that model AECI scores have been growing 1.9 times as fast since October 2025 as before, though the relationship between benchmark scores and real-world capabilities remains unclear. AA evaluated more than 100 inference endpoints and found that nearly a quarter scored below the reference deployment. Some GLM-5.3 Flash endpoints scored just 22%–29% on HLE-250, compared with 55% for the reference deployment.
Model weights and training know-how are reaching developers at different paces.
Thinking Machines has released Inkling 2 Small, a multimodal MoE with 594B total parameters and 17B active parameters, now available on Tinker. It supports image inputs and scores 31 on the Artificial Analysis Intelligence Index, a significant improvement over its predecessor. Its weights remain closed for now, with a release under Apache 2.0 planned for next week after external safety testing and red-teaming are completed.
Read original →Ant's Ling has released Ling-3.2-flash, with 560B total parameters and 25B active parameters. The model focuses on complex visual understanding, long-horizon multimodal agents, 3D understanding and interaction, and medical intelligence. Official results report a score of 92.21 on OmniDocBench1.5 and 85.86 on Vision2Web, both above the Opus 5 (max) configuration used for comparison, which scored 90.56 and 80.20, respectively.
Read original →Hugging Face has published a free MoE training handbook of roughly 70,000 words, covering expert counts and sharing, routing choices, communication optimization, and deployment. The handbook draws on roughly 800,000 GPU-hours of experiments, equivalent to about $2 million. Thomas Wolf said that 18 of the 20 best open models currently use MoE, but training know-how remains concentrated among a few teams, prompting the team to run its own experiments and document the findings.
Read original →🔍 Analysis: Inkling 2 Small activates 17B of its 594B total parameters, while Ling-3.2-flash activates 25B of 560B. These figures show that similar total parameter counts do not mean the same number of parameters are used for each inference call; model selection cannot rest on size alone. Hugging Face's handbook shares lessons on routing, communication, and deployment, filling in the engineering knowledge needed to take these models from architecture to operation. For development teams, obtaining weights and mastering efficient training and deployment remain two separate hurdles.
METR estimates a 1.9-fold increase in score growth, while spelling out the limits of that finding.
METR has released a redacted risk report using internal Anthropic data to assess whether AI is accelerating its own R&D. Since October 2025, Anthropic models' AECI scores have grown 1.9 times as fast as before, with a 95% interval of 1.5–2.8 times. METR cautions that the estimate depends on analytical choices, and the relationship between benchmark scores and real-world capabilities remains unclear.
Read original →The cost advantages of inference services need to be evaluated alongside accuracy, code validation, and runtime isolation.
AA has released its Endpoint Accuracy Index v1.1, evaluating more than 100 serverless endpoints and finding that nearly a quarter scored below the reference deployment. Some GLM-5.3 Flash endpoints had output limits of roughly 32K tokens and scored just 22%–29% on HLE-250, compared with 55% for the reference deployment. Two Muse Spark Open endpoints also produced tool-argument type errors.
Read original →CMU's ServerBench tasks coding agents with building complete inference serving stacks. Of 33 runs, 22 passed validation and 13 outperformed the strongest baseline. A system Claude Sonnet 5.5 built for Gemma E4B delivered up to 4.87 times vLLM's value per unit of GPU cost. However, systems delivered in failed runs often contained untested code paths.
Read original →Replit's Amjad Masad introduced a new sandbox design, saying its custom Linux kernel build can avoid 96% of Linux zero-day vulnerabilities. Masad said there are currently many such vulnerabilities. He emphasized that Replit relies on its own custom kernel build.
Read original →🔍 Analysis: ServerBench's results of up to 4.87 times vLLM's value per unit of GPU cost and the endpoint accuracy gaps found by AA point to the same validation problem: a cheaper, faster service does not necessarily preserve the model's full capabilities. Performance testing must also check output limits, accuracy, and tool calling. Untested code paths in failed ServerBench runs and Replit's custom kernel, which it says avoids 96% of Linux zero-day vulnerabilities, concern two different layers: application code and kernel isolation. Enterprises cannot substitute sandbox protection claims for validation of the code delivered by agents.
Gemini Omni 1.2 Flash extends clip length and supports multiple rounds of edits to the same clip.
Google has released Gemini Omni 1.2 Flash, increasing native video length from the previous generation's 10 seconds to 15 seconds. The new version improves consistency with reference material and supports editing the same clip through multiple conversational turns without regenerating the entire clip. Pricing for 720p video is 30% lower than Omni 1.1, at $0.07 per second. It is now generally available through the Gemini API and Gemini Enterprise.
Read original →Account queries and spending summaries are moving into chat, while transactions remain outside its permissions.
Claude began gradually rolling out its Personal Finance beta to Pro, Max, and Ultra users in the US this week. After connecting bank and card accounts, users can ask in natural language about balances, summarize spending, or find duplicate charges without opening each banking app to check. The feature has read-only access: it cannot transfer, manage, or invest funds, and it does not provide investment advice. There are no ads or sponsored recommendations, but answers may contain errors.
Read original →