Podscan's founder introduced their self-developed generative AI system, which includes four core modules: automatic scoring after six hours (using GPT-4o) to identify high-potential trial users; backend generation of personalized follow-up emails; automatic recommendation of the first monitoring keyword via GPT-4o mini during registration; and support for creating alerts and podcast lists using natural language within the dashboard. All features run in the background without exposing AI interfaces to users, with an average daily cost of approximately $0.20.
OpenAI released the more advanced speech-to-speech model gpt-realtime. The Realtime API added support for MCP servers, image input, and SIP phone calls.
NVIDIA has released Nemotron Post-Training Dataset v2, a multilingual reasoning dataset containing 6 million samples. It covers French, Spanish, German, Italian, and Japanese, preserving English reasoning chains while employing mechanisms such as line-level translation, format constraints, and language identification to control quality. Accompanying this release is NVIDIA Nemotron Nano 2, a 9B-parameter hybrid architecture model that supports configurable thinking budgets; the model weights have been open-sourced on Hugging Face.
DeepSeek-V3.1 is now live; deepseek-chat corresponds to its non-thinking mode, while deepseek-reasoner corresponds to its thinking mode. The new model adopts a hybrid reasoning architecture, improving thinking efficiency and Agent capabilities. It achieved a score of 66.0 on SWE-bench Verified, 54.5 on SWE-bench Multilingual, and 31.3 on Terminal-bench.
Hugging Face's official blog introduces how to connect Claude to the Hugging Face MCP Server to call image generation Spaces such as Flux.1 Krea [dev] and Qwen-Image.
Hugging Face's official blog published the "From Zero to GPU" tutorial, systematically introducing how to build production-ready CUDA kernels using the kernel-builder library.
A Hugging Face blog post introduces how the Model Context Protocol (MCP) supports scientific discovery, proposing a three-layer abstraction of manual research, scripted tools, and MCP integration. Using research-tracker-mcp as an example, it illustrates how to call resources from platforms such as arXiv, GitHub, and Hugging Face via natural language instructions; it also provides quick access paths and client configuration guidelines.
AI-MO released the Kimina-Prover-RL training pipeline along with accompanying models AI-MO/Kimina-Prover-RL-1.7B (76.63% Pass@32) and AI-MO/Kimina-Prover-RL-0.6B (71.30% Pass@32).
Hugging Face and Arm jointly released ExecuTorch 0.7 beta, which enables the KleidiAI acceleration layer by default for the first time, utilizing Arm SDOT instructions (supported since Armv8.2) to optimize Int4 matrix multiplication.
Arm has released Neural Super Sampling (NSS), a real-time temporal upscaling model designed for neural accelerators (NX) in mobile devices. It can upscale rendered frames from 540p to 1080p within 4ms while reducing GPU load by 50%. The model has been integrated into Unreal Engine via two plugins and supports the Vulkan ML extension.
Figure officially announced that its Vision Language Action (VLA) model, Helix, has achieved fully autonomous towel folding by a multi-fingered humanoid robot for the first time. This was accomplished without modifying the model architecture or training hyperparameters, relying solely on the addition of new datasets. The task requires real-time adaptation to towel deformation, fine finger control, and error recovery, without depending on explicit object representations. Helix also simultaneously learned natural interaction capabilities such as eye contact and gestures.
Hugging Face has released TextQuests, an open-source benchmark based on 25 classic Infocom text adventure games, designed to evaluate the autonomous reasoning capabilities of LLMs in long contexts without reliance on external tools.
Hugging Face released FilBench, a comprehensive evaluation benchmark designed specifically to assess large language model capabilities in Tagalog, Filipino (standard Tagalog), and Cebuano.
Hugging Face has released the open-source tool AI Sheets, which supports no-code dataset building, enrichment, and transformation. It can be deployed locally or on the Hub, integrating with thousands of open-source models on Hugging Face Hub (including OpenAI's gpt-oss) as well as local models. Users write prompts via a spreadsheet-like interface to generate new columns, support manual cell editing, use like feedback to automatically inject few-shot examples, and export to the Hub to generate reusable configuration files.
Hugging Face has introduced ND-Parallel support in the Accelerate library, allowing users to flexibly combine data parallelism (DP) via the ParallelismConfig class or Axolotl configuration files.
The Hugging Face TRL library has released several alignment capability upgrades for Vision-Language Models (VLMs), including the new Mixed Preference Optimization (MPO) algorithm.
PyTorch officially released version 2.8.0, which includes several major updates: new Inductor CUTLASS backend and Graph Partition support, and experimental Intel CPU quantized LLM inference.
OpenAI for Government announced a partnership with the U.S. General Services Administration (GSA) to provide ChatGPT Enterprise to all federal executive branch employees at near-zero cost over the next year.
OpenAI has released two open-source large language models, gpt-oss-120b and gpt-oss-20b, under the Apache 2.0 license. With parameter counts of 120B and 20B respectively, these models outperform same-sized open-source models on reasoning tasks, possess strong tool-calling capabilities, and are optimized for efficient deployment on consumer-grade hardware.
OpenAI releases the open-source large language model family GPT OSS, including gpt-oss-20b (21B parameters, 3.6B active) and gpt-oss-120b (117B parameters.
NVIDIA released the open-source model Llama-3.3-Nemotron-Super-49B-v1.5, which works with Llama 3.3-70B Instruct to form the AI-Q deep research agent stack.
The TII UAE team released the 3LM (علم) benchmark, which includes Native STEM (865 real-world Arabic secondary school STEM multiple-choice questions), Synthetic STEM (1,744 synthetic high-difficulty STEM questions), and Arabic Code Benchmarks (Arabic-translated versions of HumanEval+/MBPP+). It supports MCQ accuracy and pass@1 code generation evaluations. All datasets, evaluation scripts, and empirical results from over 40 models have been open-sourced.
Hugging Face blog published a tutorial demonstrating how to build an MCP server using Gradio, wrapping the IDM-VTON virtual try-on model as a tool callable by LLMs, and invoking it via VS Code AI Chat; includes complete Python code, mcp.json configuration examples, and usage instructions.
Mistral AI announced the launch of the Codestral 25.08 model and a complete enterprise-grade coding stack, including Codestral 25.08 (code completion) and Codestral Embed (code embedding).