← Past Issues
← KeepSanity
AI News Digest
Week 4, Oct 2026
2026-10-19 — 2026-10-25
product updates
Claude Duet: New personal assistant agent extracts insights from user contexts via third-party integrations.
models
Kimi K3.1 Pro: New models drop hallucination rates to 8.3% on Hallu Bench Hard.
Voxtral 3: Mistral's closed TTS and STT models enable zero-lag conversational audio responses.
Timestral: Mistral's new time-series foundation models enable users to deploy forecasting in minutes.
Dessa: Block's speech-only conversation model enables duplex audio interactions without taking turns.
MiniMax M3.1 Flash: New model delivers inference speeds 3 times faster than its predecessor.
top-rated papers
Spectral-Spatial Interpretation: New attribution method outperforms MC-Owen estimators using only 6% of the evaluation budget.
StreamEval: Benchmark for interactive streaming video generation reveals low request execution reliability across 37 models.
BizBench: Real-world benchmark shows Claude Opus 5 and GPT-6 Sol achieve 64.0% and 61.3% strict success.
tools
DeepInfer: Inference framework for DeepSeek V4.1 Flash serves 155 trillion tokens/day.
OpenShell: Nvidia releases a safe and private runtime for managing fleets of autonomous AI agents.
community
WorldTUI: A fine-tuned 360M model trained to simulate a fully functional terminal environment.
hardware
Cerebras: Deployment of trillion-parameter models on WSE-3T achieves 1500 tokens per second.