OpenAI hacked, Jev, Google’s RSI, Qwen 3.8 Omni, Bonsai 2, new Gemini Live: AI NEWS

Open-source model built on MiniMax H3 that changes the camera angle and timing of an existing video. It reconstructs a rough 3D point cloud with VGGT Omega, then re-renders the scene along a new camera path. The base model is about 62 GB.

R2R2T2Available in EU

Open-source real-time transcription model of about 4 GB. It shows the lowest word error rate and latency compared with other transcription models, including GPT Live Transcribe, in English and Chinese.

JAJing and DaoXGen LabsUnverified

Generative world simulation split into Dao, a persistent world-state engine, and Jing, a per-character experience model. It supports multiple players in one consistent environment and open-ended generation. It is a research preview with code released, fine-tuned from MiniMax H3.

DRDream RSIGoogleUnverified

Google project in which an agent improves its search strategy by learning from discovery trees of its past searches, without changing the model weights. It outperforms other recursive-improvement methods on mathematical optimization and matches GPU kernel results with over 2x fewer generations.

G3Gemini 3.8 LiveGoogleUnverified

Google's real-time voice model, with an Extended Thinking variant for deeper reasoning. It can call tools while still speaking, auto-detects 97 languages and accepts live visual input. It is many times cheaper than GPT Live and Grok Voice, and is available in the Gemini app, AI Studio and Search Live.

Q3Qwen 3.8 Omni FlashAlibabaUnverified

Fast multimodal model that takes text, audio, image and video input, with a 1M token context window and real-time streaming. It beats Gemini 3.8 Flash on many audiovisual benchmarks at much lower cost. It is available via the Qwen platform API and is not open source.

Q3Qwen 3.8 Live TranslateAlibabaUnverified

Real-time AI interpreter that translates live audio or video, separates and labels speakers, and clones voices. It supports 60 input languages and 29 spoken output languages. It leads competitors on win rate, latency and error rate, and is available via API with an online demo.

N3Needle 3Cactus ComputeUnverified

Tiny model of 8-29 MB (29M-121M parameters) built on a simple attention network, designed for phones, watches and microcontrollers. Its 'intelligence laddering' lets one set of weights run with 2 to 20 layers. It beats models 10x its size on mobile tool-calling benchmarks.

G5GLM 5.3 FlashZhipu AIAvailable in EU

Z.ai blog describing an AI infrastructure agent powered by GLM 5.3 that optimized and deployed GLM 5.3 Flash across a cluster of over 100,000 accelerators in under 2 weeks. It used 'dense feedback' from logs, traces and profiling, and sped up a decode kernel by 1.71x. Humans still set goals and safety boundaries.

MVMimo V2.6XiaomiUnverified

Xiaomi publicly streams the reinforcement learning run of its upcoming MiMo V2.6 model in a live dashboard. It shows cost over $1.6M, about 3B tokens per step and benchmark gains such as Deep Sweep rising from 58 to 67.

GPGPT PolicyUnverified

Framework for in-context robot learning in which a vision-language model lets a robot learn a task from a human video, a goal image or feedback on failed attempts, without retraining. Real-world tests went from failure to 2 of 3 successes on tasks like picking up a towel and unscrewing a bottle cap. Code is released.

GAGPT-6 AstraOpenAIUnverified

GPT-6 Astra helped decrypt an 82-character German WWII Enigma message that had been unsolved for about 80 years. It searched archives, built its own Enigma simulator and ran experiments in parallel for roughly 10 hours, with a human choosing the target and guiding the work.

JEJevType Safe AIUnverified

A 'system one' model that returns calibrated confidence scores over predefined answer options in about 70-500 ms, so it cannot invent answers (the claimed 0% hallucination). It is meant for fast structured decisions between agents and software, and can even play Doom. It is API-only with a waitlist.

LALayaAvailable in EU

Open-source (Apache 2) system-one model similar to Jev that gives confidence scores over multiple-choice options. Its authors say they proposed the concept first, and self-reported benchmarks show it slightly ahead of Jev. It comes in base, multilingual and long-context versions; the main model is 2.37 GB and fine-tuning scripts are included.

BNBespoke Nimble 9BBespoke LabsAvailable in EU

Open-source Jev-like model: a 193 MB LoRA adapter on Qwen 3.5 9B that outputs confidence scores for given options. It performs close to Jev.

AFAstra for LawOpenAILimited in EU

GPT-6 Astra configured for professional legal work with a dedicated legal search index covering case law, regulations and court rules. It can research authorities, draft documents and connect to law-firm tools, with zero data retention for eligible customers.

Three Hacktron researchers broke into OpenAI's internal systems in under 72 hours using Claude. They chained a libheif image-decoder bug on the community forum with an SSO flaw to take over ChatGPT and Codex accounts, for under $3,000 in tokens. OpenAI patched it and paid a $6,500 bounty.

MCMiniMax Code CLIMiniMaxAvailable in EU

Open-sourced coding harness that controls how a model reads files, calls tools, manages permissions and tracks progress. Paired with Kimi, it outperforms other harnesses on tasks while being faster and cheaper. It works with any model.

B2Bonsai 2 27BPrismMLUnverified

Ternary-weight (-1, 0, +1) compression of Qwen 3.8 27B that shrinks it from 55.6 GB to 5.9 GB, over nine times smaller, with similar benchmark performance. Versions include MLX for Apple devices and Q1 (5.95 GB) and Q2 (7.21 GB).

O1Occamy 1.0Accio LabAvailable in EU

Medium-sized open model, a post-train of a Qwen 35B MoE with 3B active parameters. It tops similar-sized models on agentic benchmarks and competes with some much larger models. Weights and training recipe are open, and the smallest GGUF is about 12 GB.

ZGZGCM-1Available in EU

Small 7B open model for mathematical reasoning and agentic search, with the full training pipeline released. The creator notes it is compared against older model generations, which is somewhat cherry-picked, though it ranks well on competitive math benchmarks.

O3Odyssey 3OdysseyUnverified

Foundation robot model that learns a world model from visual data, then uses an action decoder to control different embodiments including arms, humanoids, drones, cars and game agents. It learned to drive on Indian streets from only 20 hours of simulated data.