article
Daily AI Model Release Report · November 12, 2026
Daily AI model release report for November 12, 2026: Google launches decision model Gemini Decision 1 and genomics model AlphaGenome 2; Runway open-sources robot model Praxis-1; Baidu releases the ERNIE YiJing model matrix; Volcano Engine launches Seedance 2.5 flash — 5 items total with official images and key capabilities.
Published November 13, 2026
Date: November 12, 2026 (Beijing time) Scope: Only models released/launched/officially announced on November 12 Beijing time; US announcements made November 11 ET that landed November 12 Beijing time are included with a note. Deduplicated against the November 11 report; entries that merely update old models are excluded.
1. Gemini Decision 1

- Publisher: Google DeepMind
- Release time: November 11, 2026 (ET; officially announced in the early hours of November 12 Beijing time)
- One-line intro: Google's first "decision-specialized" model — instead of generating text token by token, it scores predefined candidate answers in a single forward pass, shipped alongside the Gemini Decisions API.
- Key capabilities:
- Scores up to 1,000 questions in a single pass across text, images, audio, video and PDFs; supports yes/no judgments, multiple choice with up to 255 options, and rubric-based scoring
- End-to-end latency of only 150–250 ms; output tokens are entirely free, input priced at $0.10 per 1M tokens — roughly 1/5 the cost of Flash-Lite
- Scores 66 on the Artificial Analysis Decision Index v2.1, ranking first among decision models on knowledge and reasoning evaluations with the lowest calibration error
2. AlphaGenome 2

- Publisher: Google DeepMind
- Release time: November 11, 2026 (local time; reported November 12 Beijing time)
- One-line intro: A new-generation DNA sequence model that pushes variant-effect prediction from "tissue averages" down to individual cell types and states, giving disease-mechanism research cell-level clues.
- Key capabilities:
- Trained on over 107 million single-cell profiles across 68 tissue types; conditional architecture enables targeted prediction for a specified cell type or state
- In tests across 28 primary immune cell types, causal variant identification rose from 16% (single-cell-fine-tuned predecessor) to 30%; direction-of-gene-expression-change accuracy reached 43%
- Analyzes unseen cell types and states without retraining (users can upload small single-cell RNA-seq datasets) and adds DNA methylation prediction; available via the AlphaGenome API for non-commercial research
3. Runway Praxis-1

- Publisher: Runway
- Release time: November 11, 2026 (ET; reported November 12 Beijing time by Cailianshe)
- One-line intro: Runway's first open-weight robot "World Action Model," transferring its video-model pretraining to physical robot control and fully open-sourced under Apache 2.0.
- Key capabilities:
- ~27B total parameters (MoE diffusion Transformer, ~2.8B activated per inference); dual-module architecture of video backbone plus action expert; each inference outputs a 1.6-second action sequence
- Pretrained on massive internet video, greatly reducing reliance on costly real-robot demonstration data; one shared backbone adapts to single-arm, dual-arm and other robot embodiments
- 46.2% task success on the NVIDIA RoboLab-120 simulation leaderboard — first among World Action Models and open robot foundation models; 88% success on 10 Franka real-arm tabletop tasks; inference service and full/LoRA fine-tuning code provided
4. ERNIE YiJing Model Matrix (Micro-expression / Video Generation / Real-time Interaction Models)

- Publisher: Baidu (Baidu YiJing)
- Release time: November 12, 2026 (Baidu World 2026 conference)
- One-line intro: Baidu YiJing announced upgrades across three self-developed video generation models, targeting the digital-human industry's two pain points — expressiveness and commercial cost — with prices down to ¥0.2/second.
- Key capabilities:
- ERNIE YiJing Micro-expression Model: AU (Action Unit)-level facial control with a quality Agent and 200+ sub-dimension modeling, presenting up to 8–9 micro-expressions per second to eliminate the stiff "AI face"
- ERNIE YiJing Video Generation Model: hybrid diffusion distillation plus sparsification for extreme inference compression, costing as little as ¥0.2/second; Fast mode generates 1:1 in real time (a 10-second video in 10 seconds)
- ERNIE YiJing Real-time Interaction Model: full-frame streaming output with sub-2-second response latency; autoregressive architecture plus retrieval-based memory supports unlimited-duration stable interaction; the "Comic-Drama Studio" enables one-click re-rendering for overseas markets
5. Seedance 2.5 flash

- Publisher: ByteDance · Volcano Engine
- Release time: November 12, 2026 (official WeChat announcement; API to launch on Volcano Ark on November 15)
- One-line intro: Volcano Engine's cost-effective video generation model — closest in quality to flagship Seedance 2.5, with doubled generation speed at a lower price, built for batch content production.
- Key capabilities:
- ~¥0.99/second for 720P text-to-video and image-to-video, about 34% cheaper than Seedance 2.5, with doubled generation speed
- Supports single clips up to 30 seconds; up to 10 images, 5 video segments and 5 audio segments per input; 480P and 720P resolution options
- Overall quality beats the previous flagship Seedance 2.0, with strong instruction following, physical motion and audio; rolling out to Jimeng, LibTV, Updream, Creao AI, TapNow and other platforms
Verified exclusions (not released today, or duplicates of items already covered in this series)
- ERNIE 5.5 Flash Preview: released November 11 and already covered in yesterday's report; today's conference was its formal showcase, so it is not counted twice.
- Qwen4 / Qwen4-Max Preview: released November 11, already covered.
- Gemini Omni 1.1 Flash: released August 27, 2026; today's coverage is recycled news.
- Tencent Hunyuan Hy4 preview: open-sourced August 28, 2026.
- Midjourney V8.2 Edit Model: launched July 2026.
- NVIDIA Nemotron 3.5 family (Super VL/Bolt/Transcribe): released October 15, 2026.
- Liquid AI LFM2-8B-A1B: released October 7, 2025; November 12 was only an ONNX-community format update.
- OpenBMB VoxCPM2: an older model released April–May 2026; only weights were refreshed on November 12.
- Sophon-Realtime-32B series: v1.1.0 dated November 11, 2026.
- Ant Group × CICC Ling-3.0-flash-Fin: seen only in aggregated AI digests; release date unverifiable, so excluded per the no-fabrication rule.
- Ciyuansuo (Token Exchange): Baidu's model trading platform — a platform/service, not a model.
