Google DeepMind has released SynthID Bio, bringing watermarking technology to synthetic biology. It embeds imperceptible signatures in biological code so that watermarks can be verified in both digital models and the physical proteins synthesised from them, while preserving their biological function.
Hugging Face launched Open TTS Leaderboard to evaluate open-source TTS models using objective metrics, reducing evaluation from weeks of voting to hours. It uses Qwen3 ASR for WER/CER, measures RTFx and TTFA on H200 GPUs and calculates speaker similarity (SIM) with WavLM embeddings, supporting multilingual and voice-cloning comparisons.
DeepSeek and Huawei have collaborated on programming tools for Ascend chips and released them all as open source. At their core is TileLang, a language that is easier to learn than CUDA. It was originally developed by researchers at Peking University, and DeepSeek has used it for about a year. The companies have also optimised a supernode comprising 128 Ascend 950 chips, with Huawei saying it provided full support. DeepSeek argues that an independent AI chip software ecosystem first needs a general-purpose language that is easy to program and can fully exploit the hardware's performance.
Anthropic published an evaluation saying Zhipu's open-weight GLM-5.3 model approaches its Claude Mythos Preview in exploit development.
OpenAI announced persistent Dots agents, GPT-6.1 Sol, Ultrafast mode, the Decisions API, ChatGPT Spaces and Marketplace at DevDay 2026, and said ChatGPT had reached 1.2 billion weekly active users.
According to sources, AI inference infrastructure provider Modal Labs is close to completing a $750 million round led by Accel at a $15.75 billion valuation. The size of the round had not previously been reported.
NVIDIA released Kumo Tabular, an open-source tabular foundation model that predicts labels for new rows in a single forward pass given labelled rows, without training, tuning or feature engineering. It supports classification and regression, offers three sizes from 28M to 215M, and was pretrained solely on artificially generated tables. It uses the commercially usable OpenMDW-1.1 licence and ranks first on TabArena, BeyondArena, TALENT and ScoringBench.
AMD announced an all-stock acquisition of AI research lab World Labs, co-founded by Fei-Fei Li, for approximately $8.2 billion, with completion expected by year-end. Founded in 2024, World Labs reached a $1 billion valuation within months and launched its first commercial product, Marble, in 2025, generating interactive 3D worlds from prompts.
AWS released a vLLM-Omni DLC for deploying Qwen3-TTS on SageMaker AI. A single persistent bidirectional connection streams text in and audio out, with playback beginning before generation finishes.
Mistral opened a hub in Munich with a research team focused on Physics AI and industrial AI, and plans to build one gigawatt of European computing capacity by 2030. It will work with BMW on crash simulation and engineering AI, Siemens Energy on industrial AI applications, and the Technical University of Munich (TUM) on automotive aerodynamics digital twins using wind-tunnel facilities.
This issue of Import AI covers several developments: Michael Levin proposes that minds are patterns from Platonic space interfacing through bodies and machines, arguing the hypothesis can be studied empirically. Stanford researchers Perry Dong and Chelsea Finn say robot pre-training has scaled, but a stable post-training recipe like that of language models is missing, mentioning their EXPO(-FT) algorithm.
H released the Holo4 agent model family, comprising 27B dense and 35B-A3B MoE versions. Both are available through H Models API, with weights open-sourced on Hugging Face in BF16, FP8, NVFP4 and 4-bit GGUF formats.
AWS presents an architecture combining EFA and DeepEP on Amazon EKS to scale MoE reinforcement learning training, increasing throughput by 40%.
Qwen3-TTS-12Hz-1.7B-Base, open-sourced by Alibaba Cloud's Qwen team, is available on Amazon SageMaker JumpStart. It deploys to fully managed real-time inference endpoints and clones voices from a few seconds of reference audio and matching text without retraining.
GitHub Security Lab released Fuzzing Taskflow, an autonomous fuzzing pipeline for C/C++ projects. Given a GitHub repository, it identifies entry points, analyses build systems, writes harnesses, runs AFL++, reads coverage reports and improves harnesses, then classifies every crash and generates vulnerability reports.
AWS released the WhisperX Deep Learning Container, adding batch inference, word-level timestamps through wav2vec2 forced alignment and speaker diarisation to OpenAI Whisper. It can be deployed to Amazon SageMaker AI real-time or asynchronous endpoints without building a custom image.
Liquid AI released experimental draft model LFM2.5-VL-DSpark for vision-language model LFM2.5-VL-3B. Speculative decoding speeds it up without changing output quality: decoding improves by up to 3.13x on devices and 2.66x on H100, with end-to-end gains of up to 2.62x and 2.27x respectively.
NVIDIA joined a global research consortium including Google DeepMind and EMBL-EBI to release predicted 3D protein complex structures for over 2,800 viruses through the AlphaFold Database, free for any scientist to use.
Microsoft Research systematically measured mobile manipulation robot workloads and found that offloading physical AI inference from onboard GPUs to edge or cloud GPUs can improve task success, support larger models and extend battery life.
NVIDIA and partners showcased AI deployment progress in Southeast Asia at Singapore AI Day. Singapore's HTX is researching public-safety AI based on Nemotron 3 Super and Nemotron 3 Nano Omni.
NVIDIA released Isaac ROS 5.0 at ROSCon in Toronto, Canada. The collection of ROS-based GPU-accelerated packages adds agentic workflows and platform support to help developers build, customise and deploy robot applications faster.
Xiaomi released the MiMo-V2.6 family, including omnimodal models MiMo-V2.6-Pro and MiMo-V2.6-Flash, plus MiMo-V2.6-Pro-UltraSpeed with up to 20 times faster output.
oMLX creator and maintainer Jun Kim joined Hugging Face full-time to support MLX. oMLX moves from a side project to a funded long-term effort, remaining Apache 2.0 open-source under Jun's leadership to improve stability and accelerate development.
Hugging Face added GGUF support to transformers. Users can choose a quantised checkpoint from the Hub and pass gguf_file to from_pretrained for local generation without additional configuration.
NVIDIA held an AI ecosystem event at the Grand Egyptian Museum, with Egyptian Deep Learning Institute participation growing over tenfold in a year. Hassan Allam received a data centre licence from Egypt's National Telecom Regulatory Authority and is developing a new centre with A15 with expected investment of $400 million. NVIDIA says four African AI factories have been announced or launched, with another 656 MW under construction.
Latent Space's AINews rounds up 15–16 September developments, highlighting two reality checks for coding-agent optimism. Steve Yegge shut down Gas Town and admitted that despite spending thousands of dollars monthly on coding-agent subscriptions, Gas Town was the only project he had built with them.
Mistral announced a partnership with Mozilla. Firefox's AI browsing assistant Smart Window (beta) is now powered by Mistral models, initially for France and North America, with the UK and Germany expected later this year.
A University of Manchester team used NVIDIA Earth-2's generative Earth-2 CorrDiff model and one year of hourly UK pollution data to train a nationwide air-pollution model at 2–3-square-kilometre resolution, completing training in two days on one eight-GPU Isambard-AI node.
Good Start Labs turned railway board game 1830 into a training environment and tested a 30B model. Both single-turn Q&A and multi-turn terminal-agent training improved game performance, but only the terminal-agent design improved Finance-Agent benchmark results.
IBM researchers added consistency guidelines to ALTK-Evolve, using Consistency Analyzer to diagnose decision points in agent trajectories that are prone to flipping.
Mistral and Cloudera partnered to integrate Mistral models into Cloudera's hybrid data platform. Enterprises can run inference in private or public clouds, on-premises and fully air-gapped environments with full control. They can train custom models on proprietary data in controlled environments while retaining ownership of data and resulting intelligence. Cloudera runs 30 exabytes of customer-managed data.
TRL v1.14's AsyncGRPOTrainer now supports training and synchronising only LoRA adapters to vLLM. A rank-1 adapter is just a few MB and can transfer through a Storage Bucket mounted to each job, without NCCL.
The Hugging Face team rebuilt most AUTOMATIC1111 functionality as the Workflow1111 canvas using Gradio Workflow. Its 11 media pipelines and 73 nodes cover text-to-image, high-resolution fixes, image-to-image, prompt matrices, VLM reverse prompting, detection-generated inpainting masks, ControlNet-style annotators, background removal, PNG Info and image-to-video.
Google DeepMind launched AlphaGenome Atlas, a platform containing predicted effects for 9 billion single-nucleotide variants across the human genome. At 1 PB, it is over 30 times the size of the AlphaFold Database.
Mistral announced a €3 billion Series D round at a post-money valuation exceeding €21 billion, led by Samsung Electronics and co-led by Scaleup Europe Fund and PSG Equity. It called this the largest equity financing by a European technology company, funding frontier research, computing, infrastructure and commercial growth. Mistral operates in 20 countries and serves more than 125 global enterprises, including Airbus, ASML and HSBC.
Google Research partnered with HHMI Janelia, the University of Cambridge and others to publish in Cell a complete connectome of a male fruit fly’s brain and central nervous system. It contains over 166,000 neurons and 125 million synaptic connections, making it the largest brain map to date by neuron count.
Hugging Face released funes, an open-source persistent memory layer for Claude Code, Codex, pi, Hermes and other coding agents. It indexes existing local sessions and performs embedding and reranking locally by default.
A Hugging Face blogger reproduced Surya Narreddi's idea of having a language model paint watercolours using TRL and OpenEnv. The model paints through around 150 lines of JavaScript with p5.brush, with datasets, RL environments, training scripts and models all open-source.
Hugging Face’s WebAI team released @huggingface/kernels, a lightweight library for loading and running optimised WebGPU kernels from the Hugging Face Hub. An initial collection of 207 kernels is also available as separate repositories under Apache-2.0.
Google Research released TimesFM-3, its first time-series foundation model pretrained natively for multivariate forecasting, with 330 million parameters and over 1 trillion pretraining time points.