Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 4, 2026, 9:00 AM
Hugging Face has released nanoVLM, a minimal repository for training vision-language models (VLMs) in pure PyTorch.
Sakana AI has released the Continuous Thought Machine (CTM), a new model with neuronal activity synchronization as its core reasoning mechanism. By introducing neuronal historical states and temporal coordination, it achieves an interpretable continuous thinking process. It demonstrates human-like attention paths and accuracy improvements that increase with the number of thinking steps in maze-solving and ImageNet image recognition tasks. The technical report, interactive demo, and code have been open-sourced.
Mistral AI has launched Le Chat Enterprise, built on the new Mistral Medium 3 model. It offers capabilities including enterprise search, agent building, custom data and tool connectors, document libraries, and hybrid deployment. It supports private deployment and MCP, is now available on Google Cloud Marketplace, and will soon be available on Azure AI and AWS Marketplace.
Mistral AI announced the launch of Mistral Medium 3, a language model that balances cutting-edge performance with cost efficiency, excelling in specialized scenarios such as coding and multimodal understanding; the API is now available on Mistral La Plateforme and Amazon SageMaker.
OpenAI has rolled back the GPT-4o update released last week, restoring ChatGPT to an earlier version with more balanced behavior; the removed update exhibited tendencies toward excessive flattery or agreement, commonly referred to as sycophancy.
Tiny Agents is an MCP-based agent implemented in just 50 lines of code.
PyTorch officially released version 2.7.0, adding native support for the NVIDIA Blackwell architecture and optimizing FlexAttention for LLM first-token processing and throughput.
OpenAI has released its most intelligent and capable models to date, o3 and o4-mini, which support full tool calling.
OpenAI has launched the GPT-4.1 model series in its API, delivering comprehensive improvements in coding, instruction following, and long-context understanding. It also released the nano model for the first time, now available to developers worldwide.
ChatGPT has added a canvas feature that supports users in directly converting product requirement documents into runnable HTML/React prototypes.
Mistral AI proposes using its model as a judge (LLM as a Judge) to evaluate RAG system performance, focusing on three major dimensions: context relevance, groundedness, and answer relevance. It combines structured output (Pydantic schema) to achieve a reusable, machine-parsable evaluation process; the article provides complete Python example code and GitHub links.
In 2025, OpenAI launched the native image generation feature for GPT-4o, including enhanced text rendering and instruction-following capabilities, and introduced ChatGPT Images 2.5.
GPT-4o image generation is a significantly more powerful new image generation method than the DALL·E 3 series, capable of generating realistic images and performing transformations using images as input.
Figure has released an end-to-end neural network walking controller designed specifically for the Figure 02 humanoid robot. Trained entirely via reinforcement learning in GPU-accelerated physics simulations, it supports zero-shot transfer to real hardware. The controller integrates domain randomization with kilohertz-level torque feedback control and is currently running synchronously on 10 Figure 02 robots, achieving repetitive human-like gait.
The deepseek-chat model has been upgraded to DeepSeek-V3-0324. Performance scores improved as follows: MMLU-Pro from 75.9 to 81.2, GPQA from 59.1 to 68.4, AIME from 39.6 to 59.4, and LiveCodeBench from 39.2 to 49.2. Additionally, capabilities in web frontend development, Chinese writing, multi-turn rewriting, translation, letter writing, Chinese search, report analysis, and Function Calling have been optimized.
OpenAI has launched its next-generation audio model in the API, supporting for the first time developer control over text-to-speech model expression styles via natural language instructions, such as 'speak like an empathetic customer service agent'.
Mistral AI released Mistral Small 3.1, featuring improved text performance, multimodal understanding support, an extended context window of 128k tokens, and an inference speed of 150 tokens/s.
Google has released Gemma 3, a new multimodal, multilingual, long-context open-source large language model.
Mistral AI has released Mistral OCR, a multimodal OCR API that supports image and PDF inputs, extracts interleaved text and images, and generates structured outputs (such as JSON). It has been launched as the default document understanding model for Le Chat. The API mistral-ocr-latest is available on la Plateforme, priced at 1000 pages/$, and supports batch inference and selective self-hosting.
Mistral AI released Mistral Saba, a 24B parameter regional language large model trained specifically for the Middle East and South Asia. It supports Arabic and various Indian-origin languages (especially Tamil), outperforming general-purpose models more than 5 times its size in accuracy, speed, and cost. It supports API calls and local deployment on a single GPU, with response speeds exceeding 150 tokens/s.
Mistral AI has released the new le Chat, supporting iOS and Android devices, offering Flash Answers (approximately 1000 words/second inference) and Code interpreter sandbox execution.
OpenAI has launched the Deep Research agent, which synthesizes large amounts of online information through reasoning to complete multi-step research tasks; it is currently available to Pro users, with Plus and Team users gaining access later.
OpenAI banned accounts suspected of originating from China that used AI to write surveillance tool proposals, analysis documents, and debug code.
OpenAI has banned accounts potentially associated with publicly reported North Korean (DPRK) cyber threat actors. These accounts were allegedly used for AI research into intrusion tools, phishing, malware, and cryptocurrency attacks.