A new method called DelTA reshapes how reinforcement learning updates propagate to individual tokens during LLM training, boosting math reasoning scores by over 3 points on top baselines. Engineers building reasoning-focused models now have a concrete technique to reduce noise from high-frequency formatting tokens polluting gradient updates.

Nanobot v0.2.0 ships goal persistence via a new /goal system that keeps an active objective pinned in Runtime Context across compaction and long tool chains. The release also bundles the WebUI inside the pip wheel and refactors the agent loop into a functional state machine with five new model providers.

vLLM v0.21.0 lands KV offloading with Hybrid Memory Allocator integration, speculative decoding support for reasoning budgets, and a C++20 build requirement that will break existing setups. Here is what teams running inference infrastructure need to act on now.

Hermes Agent v0.14.0 ships a local OpenAI-compatible proxy that lets coding tools like Codex, Aider, and Cline hit Claude Pro or ChatGPT Pro without API keys. The release also cuts cold-start time by roughly 19 seconds and lands xAI Grok via SuperGrok OAuth with a 1M token context window.

Firecrawl v2.10 ships a /parse endpoint for local file ingestion, a Lockdown Mode for zero-outbound scraping, and new question and highlights formats that cut token usage by up to 100x. Go, Ruby, and PHP SDKs join the official lineup.