Mistral released the public preview of Mistral Large 4 (ML4), a natively multimodal model with 1 trillion parameters and 49 billion activated parameters, with weights to be available for download by the end of the month.
Why it matters: The official release covers parameter counts, benchmark scores, training infrastructure and the open-weights timeline, letting readers assess its real standing in the open-source camp.
Google DeepMind released EmbeddingGemma 2, a 740M-parameter model built on the Gemma 4 architecture under the Apache 2.0 license, mapping text, code, images, video and audio into a shared embedding space.
Why it matters: The official release provides concrete metrics such as parameter count, memory footprint and benchmark scores, helping readers judge whether it suits on-device retrieval scenarios.
Telecom operators are building their AI strategies on open models; in NVIDIA's latest State of AI in Telecommunications report, 89% of respondents said open-source models and software are important to their AI strategy.
OpenAI released GPT-6.1 Sol, priced at $2/$10 per million input/output tokens, ranking 5th on the Agent Arena, 39% cheaper than GPT-6 Sol with a score 1.52 points higher, and 81% cheaper than Astra.
Earendil's agent framework Pi released Pi 1.0 and Pi Durable, both of which made the HN front page.
AWS released the open-source decision model Strands Decider 2B, inspired by TypeSafe's Jev, for quickly and cheaply making choices among predefined options while providing a confidence score.
Ai2 released Olmo-core 3, a redesigned open-source MoE training system that scales to trillion-parameter sizes. It switches from FSDP to a DDP-based approach, with a 47B-parameter MoE reaching a throughput of 52 on 8 NVIDIA B300 GPUs.
AI startup Photon announced a $4.5 million seed round co-led by Gradient and A*, with Vercel and HongShan among the participants. The company helps developers build agents running on messaging apps like iMessage and WhatsApp, and previously held an 'app funeral' to express its view that the app era is ending.
The Firefox 157 update introduces an interface redesign on desktop and mobile, and Mozilla Firefox head Ajit Varma explained the product strategy for reaching the mass market in an interview with Ars Technica.
DeepSeek partnered with Huawei to open-source programming tools for Ascend chips, centered on TileLang, a programming language easier to use than CUDA, plus libraries for compute and inter-chip data movement; the two also optimized a supernode of 128 Ascend 950 chips.
Hugging Face has launched the Open TTS Leaderboard, evaluating open-source TTS models with objective metrics including Qwen3 ASR WER/CER, RTFx and TTFA on H200, and WavLM speaker similarity, cutting per-model evaluation time from about two weeks to hours.
Why it matters: The post explains how objective metrics shrink TTS evaluation from weeks to hours, and why mainstream arenas structurally underrepresent open-source models.
NVIDIA has released the open-source tabular foundation model Kumo Tabular, now on Hugging Face. Given labeled tables, the model predicts labels for new rows in a single forward pass with no training, tuning, or feature engineering; it supports classification and regression, is Transformer-based with column, row, and in-context attention, and was pretrained entirely on synthetically generated tables.
Why it matters: The official post covers the architecture, pretraining recipe, and four benchmark results of a zero-finetuning tabular foundation model, letting readers judge whether it can replace traditional gradient-boosted-tree workflows.
Mistral opened a new center in Munich with research teams focused on Physics AI and industrial AI. Since acquiring Emmi AI in May 2026, more than 30 physics and engineering AI researchers have joined; the company is currently working with BMW on crash simulation, with Siemens Energy on industrial AI, and with TUM on digital twin research for automotive aerodynamics.
H Company has released the Holo4 agent series, available in two sizes: 27B dense and 35B-A3B MoE, alongside Holotron4 Nano built on Nemotron 3 Nano Omni.
Why it matters: The official release of a generalist computer-use agent series details multi-interface adaptation design, OSWorld 2.0 score comparisons, and open-source trajectories, letting readers evaluate new open-source agent options.
Hugging Face announced that the Hub now supports RL Environments, which exist as dataset repositories tagged with rl-environment, with no new repository type, registry, or registration process.
The Latent Space podcast spoke with OpenRouter co-founder and CEO Alex Atallah and AMP's Anjney Midha, tracing how OpenRouter grew from the early wave of open-weight models like Llama and Alpaca into a neutral routing layer serving over 10 million developers and handling over 100 trillion tokens per day, and why it was acquired by Stripe.
Antonio Morales of GitHub Security Lab has released Fuzzing Taskflow, an autonomous fuzzing pipeline for C/C++ projects built on the Taskflow Agent — you only specify a GitHub repository, and it automatically identifies entry points, analyzes the build system, writes harnesses, runs AFL++, and triages crashes.
Why it matters: The author explains the design of the automated fuzzing pipeline, showing how LLM agents can take over manual steps like coverage analysis and crash triage.
A Microsoft Research study shows that running physical AI inference solely on robots' onboard GPUs limits performance, battery life and scalability, and offloading inference to edge or cloud GPUs can significantly improve it.
NVIDIA held AI Day Singapore on September 22-23 at the Raffles City Convention Centre, where it and its partners showcased AI advancements across the Southeast Asia region.
Simon Willison released llm-anthropic 0.29 on September 22, 2026. The original post provides no further version details or feature notes.
At ROSCon in Toronto, Canada, NVIDIA released Isaac ROS 5.0, a suite of GPU-accelerated open source packages built on ROS, adding agentic workflows to help developers and AI agents collaborate on building robotics applications.
Xiaomi released the natively omni-modal MiMo-V2.6 series; MiMo-V2.6-Pro has 1.02T total parameters with 42B active, is open-sourced under the MIT license, and tops Artificial Analysis's open weights leaderboard with an Intelligence Index score of 46.
Hugging Face has announced that transformers can load GGUF models — passing a gguf_file to from_pretrained lets you run quantized checkpoints from the Hub on your local machine.
Why it matters: Hugging Face provides official usage of GGUF in transformers plus benchmarks against llama.cpp, helping readers pick a viable local inference path.
Jun Kim, creator and maintainer of oMLX, has joined Hugging Face. oMLX will shift from a side project to one with full-time maintenance and funding, while remaining Apache 2.0 open source and led by Jun. One team focus is simplifying the pipeline from transformers model definitions to reference MLX implementations, so new transformers models run on MLX faster.
The UK AI Security Institute (AISI) has begun using EvalEval's infrastructure to publicly share evaluation results, including verified results for five benchmarks—HealthBench, FrontierMath, Humanity's Last Exam, SWE-Bench Pro and Terminal-Bench 2.0—as well as two cyber evaluations, Cyber CTFs and The Last Ones.
Mistral announced a partnership with Mozilla: the Firefox Smart Window (beta) AI browsing assistant is now powered by Mistral models, serving users in France and North America, with the UK and Germany expected to follow later this year. The feature helps users understand complex searches, recover important content clicked away, and organize information based on browser tabs. Conversations are not stored on Mozilla servers by default, and Mistral pledges zero data retention.
Hugging Face released Workflow1111, rebuilding AUTOMATIC1111's main features as a single Gradio Workflow canvas with 11 media pipelines and 73 nodes, covering text-to-image, hi-res fix, img2img, prompt matrix, VLM interrogation, detection-generated inpaint masks, ControlNet-style preprocessors, background removal, PNG Info, and image-to-video.
Mistral announced a €3B Series D round at a post-money valuation above €21B, the largest equity round by a European tech company in the three years since its founding.
Why it matters: The official announcement fully discloses the round size, lead investors, and use of funds, showing the capital backing and expansion direction of Mistral's sovereign AI strategy.
Google Research, together with HHMI Janelia, the MRC Laboratory of Molecular Biology, the University of Cambridge and others, has published in Cell a complete connectome map of the male fruit fly brain and central nervous system, containing over 166,000 neurons and 125 million synaptic connections — the largest brain map to date by neuron count.
Why it matters: The article details the connectome's scale, the AI reconstruction methods, and an open-source viewer, so readers can gauge the current limits of connectomics and the resources available.
H Company released NeoMME, a multilingual multimodal encoder in 260M and 800M sizes. A single bidirectional Transformer processes both text tokens and 32×32 image patches, trained from scratch with a masked discrete diffusion objective, with no reliance on a pretrained vision tower or a causal language model. Both versions are open-sourced under Apache 2.0.
Hugging Face's WebAI team released the @huggingface/kernels library, which loads and runs 207 optimized WebGPU kernels from the Hub, under the Apache-2.0 license.
Why it matters: The official release includes 207 WebGPU kernels, a JS loader, the crowdtesting tool Fleet, and measured comparison data against ORT WebGPU.
Google Research released TimesFM-3, a time series foundation model with 330M parameters, pretrained on over 1 trillion timepoints of real and synthetic data, and the first to natively support zero-shot multivariate forecasting.
Microsoft Research released GigaPath-Flash and GigaTIME-Flash, using knowledge distillation to compress the billion-parameter GigaPath encoder into a 22M-parameter ViT-S encoder, with performance within 3% of the original model at roughly 50x lower compute.
AMD, Saudi Arabia's Ministry of Communications and Information Technology (MCIT) and the Digital Cooperation Organization (DCO) announced the launch of an ecosystem program offering developers and startups open software tools, training and community programs. The first workshop was held at LEAP 2026, introducing the AMD AI Developer Program, the AMD ROCm open software ecosystem and the DCO's AI ethics and responsible governance training courses. Going forward, the DCO will work with AMD and member states to explore expanding AI capacity building and knowledge sharing.
Voice Arena and Hugging Face released four Monsoon evaluation sets for the Open ASR Leaderboard, covering Indian English and Hindi, with Hindi the first non-European language on the multilingual leaderboard.
Why it matters: The original details the Monsoon sets' split structure, speaker coverage and regional analysis examples, helping readers understand sources of model differences beyond aggregate WER.
AMD released the ROCm 10 software stack and announced that ROCm.AI is now generally available, an AI-native development platform comprising three components: ROCm Hyperloom, AMD Skills, and ROCm CLI.
Why it matters: AMD officially details the capabilities of ROCm.AI's three components along with 3.3x inference and 2.4x training gains, letting developers evaluate whether to migrate to the new workflow.
TrendForce analysis says that after Moonshot AI released Kimi K3 in July 2026, the gap between leading Chinese and US models narrowed significantly, with market discussion of a "second DeepSeek moment."
Sentence Transformers v6.0 adds the MultiVectorEncoder model type, and this tutorial demonstrates how to train and finetune ColBERT-style multi-vector retrieval models.
Why it matters: The author backs every training choice with measured data, teaching a multi-vector model domain finetuning method reproducible on a single GPU.
IBM released the Granite 4.2 reasoning model family, comprising 3B, 8B and 30B dense decoder models built on the Granite 4.1 base (pretrained from scratch on about 15T tokens, with a five-stage strategy extending context to 512K), post-trained with SFT and multi-stage GRPO reinforcement learning, all open-sourced under the Apache 2.0 license.
Why it matters: The official disclosure covers the full training pipeline from pretraining and SFT to multi-stage GRPO and agentic RL, transferable to similar model-building workflows.
Hugging Face introduces gr.Workflow in Gradio, turning AI application pipelines into an interface: steps are described with typed nodes, a drag-and-drop canvas is provided, each node can run, and every intermediate result is visible.
Why it matters: The original covers gr.Workflow's node types, REST API and deployment options, with multiple copyable Space examples, making it easy to start building AI pipelines.