Marin 535B-A23B open training run hits 13%

The new Cowork runs model inference and its VM in the cloud, giving each session its own sandbox instead of the locally shipped Anthropic VM. Anthropic's Felix Rieseberg cites disk, battery and performance costs, plus work stopping when a laptop closes, as reasons for the shift.

The 660M-parameter image model now runs a 4-step turbo mode that generates 7 images per second on an RTX 5090, addressing complaints that its small size didn't translate to fast generation. An earlier release post describes it as a 670M-parameter uncensored image model.

Claude Code v2.1.290 adds serverToolUses to a mod's turn.step hook result, exposing the API-run tool calls (the advisor) with id, name, input, and start/end times. Plugin hooks gain agentId on tool.check so a hook can distinguish a subagent's permission check from the main session's.

Browserbase's Miguel González Fernández and Microsoft Research's Corby Rosset present research on verifiers for computer-use and web agents, where an official judge scored agent success at 74% while a better verifier put it at 38%.

Amazon Bedrock adds GLM 5.3, an open-weight model aimed at coding and agentic workloads such as refactoring repositories spanning hundreds of files and sustaining multi-hour agentic workflows with tool use.

An APK teardown found a "Gemini Calling" intro screen with examples like "Call Mom and tell her I will be 15 minutes late" and "Call John and ask if he is coming for dinner." The teardown also surfaced granular per-app permissions, including phone call settings, though Google may never ship the expansion.