Genie 4, Runway Ads, and Nemotron 3.5 Transcribe Lead a Busy Week in AI Model Releases

Key Takeaways

1. Google DeepMind released Genie 4, an interactive world model combining real-time physics, reasoning, and audio; early users report it can run DOOM in real time.
2. Runway launched Runway Ads, which turns text descriptions into brand-aligned video ads and learns from performance to improve later versions.
3. NVIDIA released Nemotron 3.5 Transcribe, an open-weight 30B MoE model with about 4B active parameters covering 40 languages, speaker attribution, translation, and audio understanding in one checkpoint.
4. Hugging Face published the LFM3 family: six multimodal models spanning text, audio, vision, and omni, with the 1.7B model supporting 256k context.
5. Mistral released a public preview of its Responses API for agents and reported that Voxtral Mini Transcribe 3 ranked second on Voice Arena's diarization plus ASR leaderboard.
6. Anthropic expanded its GSA OneGov agreement to offer Claude at a 40% discount to U.S. government users and disclosed early drug-discovery work.
7. Perplexity shipped a new voice mode to Pro and Max users and demonstrated a GTA V project built through its Computer MCP bridge.

1. World Models and Creative Tools

1. Demis Hassabis announced Genie 4, describing it as integrating real-time physics, reasoning, and audio and hinting at "simulations within simulations." A quoted post called it the most realistic and interactive world model to date, explorable in real time with responsive scenes, sound, and character motion and speech 1. Ethan Mollick reported that Genie 4 can run DOOM in real time, with every computer and wall in an office scene running the game, and said that despite glitches it is impossibly advanced compared with a few years ago 2 3.
2. Runway launched Runway Ads, which lets users describe a creative direction in text and generate brand-aligned video ads adapted to different platforms. The system learns which ads perform better and generates improved versions accordingly; Runway said teams including Lovable have used it to scale ad creative and shorten production time 1.

2. Open Models and Speech

1. NVIDIA released Nemotron 3.5 Transcribe, an open-weight, promptable 30B MoE model with roughly 4B active parameters. It combines 40-language transcription, speaker attribution, translation, contextual prompting, and audio understanding in a single checkpoint 1 2.
2. Hugging Face published the LFM3 series, built on a tri-hybrid architecture with a multimodal encoder across six model instances covering text, audio, vision, and omni, including generation and decision models. LFM3-1.7B supports 256k context, LFM3-VL-2B accepts video input, LFM3-Audio-2B supports tool calling, voice chat, and reasoning in 18 languages, and LFM3-Omni-2B fuses vision, audio, and language; d1-2B and d1-Omni-2B are decision models that output typed results and calibrated probabilities in a single pass 1. The posts said LFM3 can be fine-tuned on a single GPU, is optimized for 14 processor classes across GPU, CPU, and NPU, and that LFM3-1.7B ranks first on the on-device intelligence Pareto frontier 2.
3. Mistral said Voxtral Mini Transcribe 3 ranked second of nine models on Voice Arena's Diarization + ASR leaderboard with 40.4% DER, ahead of the larger Voxtral Small Transcribe 3 at 41.1% DER and fourth place. The test used 139 held-out conversations, human-verified references, a strict 0 ms collar, and no preset speaker count or audio cleanup; Mini was slightly better in face-to-face conversation (40.6% vs 41.5%) while Small was slightly better on online calls (39.1% vs 39.5%), and overlapping speech was the biggest challenge, with missed speech accounting for 82% of Mini's total DER 1. Mistral also launched a public preview of its Responses API for building agents, returning model answers, reasoning, and tool calls as separate entries and supporting all Mistral-hosted chat models including Large 4 2.

3. Enterprise, Government, and Agent Products

1. Anthropic said its scientists have spent months building deep research collaborations with six new drug-discovery organizations and have started their own drug-discovery work, publishing early results and an overall life-sciences plan 1. It also expanded its partnership with the U.S. General Services Administration through a new OneGov agreement offering Claude at a 40% discount to federal, state, local, tribal, and territorial governments 2.
2. Perplexity announced a new voice mode for all Pro and Max users on web and the Mac app, allowing voice conversation with the computer while advancing multiple tasks in parallel 1. It also described an experimental project in which Perplexity Computer designed a GTA V mission: it consulted FiveM documentation, built the mission as a FiveM resource in Lua, and added an MCP bridge where a local Node service logs in to the Perplexity Computer MCP server via OAuth and turns in-game prompts into real Computer tasks, with the in-game phone streaming Computer steps during a chase 2. For voice, Computer rented cloud GPUs to run Qwen3-TTS locally rather than using a TTS API 3.

4. Evaluation and Safety Signals

1. Hamel Husain said he found defects in about 65% of TAU-banking tasks, mainly policy not being correctly executed: models can take many out-of-bounds actions without penalty, lie directly to customers without penalty, and tool-call checks are coarse, with failed calls having no consequence. He grouped the defects into three categories: unpenalized out-of-bounds actions, unpenalized lying to customers, and tool-call checks lacking field and follow-up validation, and said some scoring directions may be correct but need hardening before use 1. He plans to review all public benchmarks because some people use flawed benchmarks to decide what data to buy and data companies produce problematic data to climb benchmark scores; he stressed this is not a rejection of Sierra and the TAU series' contributions to tool-calling evaluation 2.
2. Ethan Mollick cited an experiment in which experts working with an AI agent followed the agent's inclination 86% of the time even when they believed they were judging independently, against 50% if there were no influence, and advised deliberately choosing when to use AI if independent judgment matters 1. He also disputed the claim that AI companies talk about existential risk to hype and lift stock prices, arguing that telling users a product could kill everyone does not promote adoption or reduce regulatory risk 2. Separately, Artificial Analysis reported that after switching to a closed benchmark, the open-weight model gap widened: leading open-weight model Kimi K3.1 Pro, whose weights are not released, scored 46, 12 points behind the leader versus a 3-point gap at v4.3, while the highest released-weight models, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max), scored 32, below Claude Haiku 5.5 (max); the gap was largest on HLE-Diamond and Terminal-Bench-Science, where Kimi K3.1 Pro scored 37% and 34% against 68% for Claude Opus 5.5 (max) and 63% for GPT-6 Astra (max) 3. The author cautioned against overreading the result but said it matches the general sense that no Fable-class open model has appeared 4.

5. Infrastructure and Applied AI

1. NVIDIA said AXA researchers used its Earth-2 open global weather model FourCastNet 3 to generate more than 1,000 simulated cyclone seasons to study global storm activity, filling a risk-assessment gap that decades of historical data cannot cover. The simulations helped understand cyclone tracks, frequency, and same-season effects on multiple regions, and the insights can feed AXA's risk assessment and targeted customer prevention 1. NVIDIA also published post-training content for Europe under the Nemotron name and discussed latency and throughput trade-offs when serving world models with NVIDIA Cosmos 3 Super on Nebius 2 3.
2. OpenAI said the Yang-Mills working group will pursue scientific collaboration to interpret AI results related to quantum chromodynamics existence and Standard Model completion, including a subproblem of proving the existence of quantum Yang-Mills theory 1.
3. SpaceX argued in comments to the FAA that rockets are no longer the bottleneck and that the FAA's outdated air-traffic process is; it said its self-built Cleared air-traffic tool is already provided free to most major U.S. airlines and that Starbase Louisiana is designed for about 30 launches and reentries per day, with Elon Musk adding that "10,000 Starship launches a year is just from Starbase Louisiana" 1. Separately, a reposted United Airlines CEO account said the airline had planned Starlink on only 65% of its fleet but decided to roll it out fully after a passenger said they switched to a regional flight to get Starlink 2.

Note: A repost cited by Elon Musk claimed a Ukraine-Russia ceasefire took effect at 4:00 Moscow time on October 27, 2026, with an expanded scope for humanitarian convoys and confirmation by Russia's chief of general staff; this remains disputed/unverified 1.