Cactus Compute released Whistle, an Apache-2.0 16.9 MB, 55M-parameter on-device speech recognizer for seven languages, designed to run locally on CPUs across phones, wearables, robots and microcontrollers. The agentic angle is unusually direct: its Needle runtime can pass audio through transcription into structured tool calls in one step, so a voice interface can control a local agent without sending audio to a cloud service. Cactus reports 11.1 ms time-to-first-token on an Apple M4 Pro and better WER than Whisper base on several listed sets, but those benchmarks are vendor-measured and Whisper leads on others. The Reddit discussion surfaced today; independent testing remains limited. Sources: https://www.cactuscompute.com/blog/whistle ; https://huggingface.co/Cactus-Compute/whistle ; https://www.reddit.com/r/LocalLLaMA/comments/1wyemcb/whistle_speech_to_text_in_a_169mb_file/
Newly surfaced research, DeReAct, separates a ReAct agent’s action proposal from two explicit gates: a Critic validates actions before execution, while a Context Manager reconstructs environment-supported state and verifies task completion. The Oct. 1 paper reports gains on GAIA and SWE-bench Verified, largest with weaker base models (6.5–7.0 points for Qwen3-Coder-480B; 4.2–5.2 for Claude Sonnet 4.5), with smaller gains as the base model strengthens; these are paper-reported results, and gating helps only when the targeted failures are prevalent and the gate is capable enough. Source: https://arxiv.org/abs/2610.02351