Simon Willison has updated llm-echo, a debugging plugin for his LLM tool that provides a fake model which simply repeats the prompts it receives. Version 0.5 makes its simulated token counts account for tool calls, system prompts and reasoning tokens, which helps when testing changes ahead of a new LLM release.
The new OpenClaw release makes its update command finish and recover on its own across devices, file systems and platforms, instead of leaving a broken install. Interrupted turns, subagent work and deferred tasks now resume after a Gateway restart, so conversations are no longer lost.
NVIDIA has published its entire Run:ai GPU orchestration platform as open source under the Apache 2.0 license. The release includes the KAI Scheduler, GPU fractioning with CUDA 13.4, quota management and a local development environment based on Kind.
Google Research released Harness2 under the Apache 2.0 license, a framework that rewrites an agent's harness at test time. A single refinement step raised GLM-5's score on WorkBuddy-Bench by 11.5 points without touching the model's weights.
A MarkTechPost tutorial walks through NVIDIA's new Structured Data Models library, which exposes several pretrained foundation models for tables through a single PyTorch interface. It bundles TabICLv2 and KumoTabular for single tables and KumoRelational for full relational databases, all predicting through in-context learning rather than training.
Kyutai, working with Inria and École des Ponts, has released REKI, a 2.4-billion-parameter vision-language-action model for robotic manipulation, trained solely on public data and equipped with a low-cost memory of recent events. On three out-of-distribution benchmarks, only models roughly twice its size outperform it, and its weights are publicly available.