Gemini Omni 1.1 Flash
Google updated its video-generation model to Gemini Omni 1.1 Flash, adding four production-oriented controls: scene extension (10-second increments up to 40 seconds total, using up to 10 seconds of prior context for consistency), keyframe control (specify first/last frames for guided transitions), a 360p draft mode (~60% faster and a third of the cost of 720p, for rapid iteration), and 4K upscaling for final output. It's live now in Google AI Studio and the Gemini Enterprise Agent Platform, and rolling to Google Flow (Plus/Pro/Ultra) and the Gemini app (scene extension only, Plus/Pro/Ultra).
Sources & depth
Anthropic opens a research preview of the Model Hardware Standard
Anthropic launched a research preview of the Model Hardware Standard (MHS), a shared driver/protocol spec that lets AI agents operate physical lab and manufacturing instruments through one common interface instead of a bespoke integration per device. The pitch: integration work that currently takes labs weeks or months drops to hours, enabling multi-device orchestration, closed-loop experiments, and autonomous error recovery without a person in the loop. It's an invite/waitlist preview for scientific research labs and advanced manufacturers ahead of an eventual open-source release; early partners named include Genentech, Carnegie Mellon, the University of Washington, and QuEra Computing.
Sources & depth
DeepMind pilots double-blind AI evaluations
Google DeepMind ran what it says is the first double-blind external evaluation of a frontier-class model. Using Confidential Computing (Google Cloud's Confidential Space), the pilot tested a Gemini Flash Lite variant such that the evaluator never sees the model weights and Google never sees the evaluator's test prompts — addressing the standing dilemma in independent model auditing, where external evaluators either risk leaking benchmark questions to the model maker or the maker risks leaking weights to the evaluator. Partners on the pilot: the Singapore AI Safety Institute, OpenMined, AVERI, and MLCommons; DeepMind points to a separate technical report for the actual results rather than stating them in the post.
Sources & depth
Gemini 3.5 Transcribe
Google shipped Gemini 3.5 Transcribe, a new speech-to-text model it calls its most precise yet: 4.0% word-error-rate in streaming mode, 2.6% non-streaming, over 85 languages with auto-detection, up to three-speaker diarization with timestamps, custom-vocabulary support for jargon, and sub-second streaming latency — a claimed 70% latency improvement over the prior Chirp 3 model. It's in public preview via the Gemini API/AI Studio and enterprise preview on the Gemini Enterprise Agent Platform now, with a consumer rollout to Gboard, the Gemini app on macOS, and Chrome to follow. No pricing published yet.
Sources & depth
Qwen3.8-Flash-Next + GLM-5.3-Flash land in Unsloth
Unsloth added local-inference support for both of this week's open-weight flagship releases: Qwen3.8-Flash-Next now runs on 75GB RAM, and GLM-5.3-Flash on 102GB RAM+VRAM, with 5x faster RAM-offloaded inference and "infinite" repeated context compaction. This is what turns two size-class open releases from "download the weights" into something a well-equipped workstation can actually run — practically the more useful update for self-hosters than the original ship announcements.
Sources & depth
Anthropic expands scientist access
Anthropic announced a bundle of researcher-access moves: 10,000 free/discounted Claude Team seats for verified scientists over the next year (free standard seats; premium seats with 5x usage limits at $15/month, eligibility gated to a verified principal-investigator-or-equivalent role at an academic or nonprofit institution); an expansion of its AI-for-Science credit program (up to $50,000 in credits per project) beyond biology into other research fields; and, via a new US-government partnership, an access program for "Mythos-class" models aimed at life-sciences researchers, with first participants already enrolled. Claude Fable models will keep restricting professional biology/drug-development queries on dual-use safety grounds.
Sources & depth