Cloudflare ships Clef and Clef-flash — open-weight decision models
Oct 1: Workers AI hosts Cloudflare-trained decision models that return typed, probability-scored answers (Jev-API compatible) instead of free-form tokens — useful for agent routing, triage, and threat classification. Clef adds a vision encoder and 64k context; Clef-flash targets sub-50 ms median latency. Weights are Apache 2.0 on Hugging Face; Cloudflare also launched hands-on RL fine-tuning that will become a self-serve platform. Strong HN showing (~472 points).
blog.cloudflare.com·Cloudflare·HN
Microsoft MAI: streaming transcription + Voice 2.1 / Flash
Oct 1: MAI-Transcribe-2-Streaming debuts as Microsoft’s first streaming ASR — 60 languages, continuous language detection, first partials just over 100 ms, #1 accuracy on Artificial Analysis streaming (AA-WER as of Sep 28). Intro price $0.54/hour through year-end. Alongside: MAI-Voice-2.1 (23 languages / 26 locales, $22 per 1M chars) and Flash (150 ms e2e for ≤45 s audio, $15 per 1M chars). Available via Microsoft Foundry, MAI Playground, Vercel, OpenRouter (voices); LiveKit coming.
news.microsoft.com·Microsoft·Unite.AI
Tavus Griffin — first “Human Interaction Model” (research preview)
Oct 1: Full-duplex video-to-video model that perceives audio/video continuously and generates expressive speech + 720p video together (not a classic ASR→LLM→TTS cascade). Tavus reports 48% of participants thought Griffin-Lite was a real person after a one-minute call (vs ~2% for their prior stack), plus #1 on NVIDIA’s VideoFDB generation and perception tracks. Griffin-Lite is limited to trusted testers pending safety/disclosure work — not general customer GA.
tavus.io·Tavus
Fastino Labs releases GLiDE decision model via API
Oct 1: GLiDE (Generalized Lightweight Decision Engine) is positioned as an adaptive-reasoning decision model for agent action/tool/approval choices. Fastino’s own Decision Index 0.2.1 run scores it 64.81 vs Jev’s published 57.91 (+6.9), leading all five areas and 31 of 38 benchmarks. Available as fastino/GLiDE on the Fastino API — treat as vendor-reported numbers until independent leaderboard confirmation.
PR Newswire·Fastino Labs