↗TiaoTechIdeas in motion
中文
☰
Home/Stay curious. Stay in the loop.
TIAOTECH / NEWS

Stay curious. Stay in the loop.

Selected industry news from AIHOT. Follow advances in AI, products and technology.

AIHOT · Selected summaries with original sources · Nov 4, 2026, 9:00 AM

AI models · Mistral AI

Mistral Releases Mistral Small 3 (24B) Model ↗

Mistral AI has released Mistral Small 3, a 24-billion-parameter large language model licensed under Apache 2.0, focusing on low latency and local deployment; it achieves 81% accuracy on MMLU.

Technology products · PyTorch Releases

PyTorch 2.6.0 Officially Released ↗

PyTorch 2.6.0 has been officially released. Major updates include: torch.compile supports Python 3.13; a new torch.compiler.set_stance API dynamically controls compilation behavior.

AI models · DeepSeek 官方更新

DeepSeek Releases Reasoning Model DeepSeek-R1 ↗

DeepSeek has released a new model, DeepSeek-R1, which can be invoked by specifying model=deepseek-reasoner; official links for update notes and invocation guides are provided.

AI models · Mistral AI

Mistral AI Releases Codestral 25.01 Coding Model ↗

Mistral AI has released Codestral 25.01, which features a more efficient architecture and an improved tokenizer compared to Codestral-2405, achieving approximately 2x faster code generation and completion speeds.

AI models · DeepSeek 官方更新

DeepSeek-V3 Released, deepseek-chat Model Upgraded ↗

The deepseek-chat model has been upgraded to DeepSeek-V3. Continue using model=deepseek-chat when making calls; the interface remains unchanged. For detailed updates, see the official DeepSeek-V3 release page.

AI models · DeepSeek 官方更新

DeepSeek-V2.5-1210 Released ↗

The deepseek-chat model has been upgraded to DeepSeek-V2.5-1210. Mathematical capabilities improved from 74.8% to 82.8% on the MATH-500 benchmark, and coding accuracy increased from 29.2% to 34.38% on LiveCodebench (08.01 - 12.01). Chinese writing and reasoning capabilities also showed improvements in internal test sets, along with optimizations to file upload and web page summarization features.

AI products · OpenAI News

OpenAI Releases Sora Video Generation Model and Opens Access ↗

OpenAI's video generation model Sora is now live at sora.com, supporting videos up to 1080p resolution and 20 seconds in length, compatible with widescreen, vertical, and square aspect ratios; users can generate new content from text or import their own materials for extension, remixing, and blending.

Research · Hugging Face Blog

Hugging Face and LLM-jp Jointly Release Open Japanese LLM Leaderboard ↗

Hugging Face and the cross-institutional project LLM-jp jointly released the Open Japanese LLM Leaderboard. It includes over 20 Japanese datasets covering 16 tasks (including NLI, machine translation, code generation, mathematical reasoning, multi-hop question answering, etc.), uses llm-jp-eval v1.14.1 for evaluation, and is backed by vLLM v0.6.3 and Japan's high-performance computing platform mdx. The leaderboard supports automatic model deployment and evaluation, with all datasets and tools being open source.

Technology products · Playwright Releases

Playwright v1.49.0 Released ↗

Playwright v1.49.0 has been released, featuring the new expect(locator).toMatchAriaSnapshot() assertion for verifying page accessibility trees, with support for generation via Test Generator and updates using --update-snapshots; it introduces a new Chromium headless mode ('chromium' channel) that more closely mimics real browser behavior; the test runner adds tsconfig configuration, test.fail.only(), multiple globalSetup/Teardown functions, screenshot='on-first-failure', and other features; HTML reports include previous/next navigation buttons; traces support grouping; Docker images are upgraded to Node.js v22 LTS; includes Chromium 131 / Firefox 132 / WebKit 18.2.

AI products · Mistral AI

Mistral AI Releases Multi-Feature Update for le Chat (Including Canvas, Pixtral Multimodal Understanding, Flux Image Generation, and Agents) ↗

Mistral AI officially announced the launch of multiple new features for le Chat (all in beta): support for web search with citations; a new Canvas interface for collaborative creation and inline editing; integration of the Pixtral Large multimodal model for deep understanding of documents and images; connection to Black Forest Labs' Flux model for high-quality image generation; and the introduction of reusable, shareable Agent features to automate repetitive workflows. All features are currently available for free.

AI products · Mistral AI

Mistral Releases Batch API ↗

Mistral has released the Batch API, available on La Plateforme. It processes high-concurrency requests at 50% lower cost than the synchronous API, supporting use cases such as customer feedback analysis, bulk document summarization and translation, vector embeddings, and data annotation. Each workspace is limited to 1 million in-progress requests. The service will soon be available on cloud provider partner platforms.

AI products · Mistral AI

Mistral AI Releases Content Moderation API ↗

Mistral AI has released a content moderation API, sharing the same source as Le Chat's built-in moderation service. It supports two input endpoints for raw text and conversational content. The model is a multilingual LLM classifier covering 11 languages including Arabic, Chinese, and English. It identifies inappropriate content across 9 policy categories and provides protection against model-generated harms such as unqualified advice and PII leakage.

Practical tips · Hugging Face Blog

Hugging Face Releases PyCharm Integration Plugin ↗

Hugging Face officially released a PyCharm integration plugin that supports right-click model insertion (e.g., microsoft/Phi-3.5-vision-instruct), real-time model card viewing, and local cache management; this feature is exclusive to PyCharm Professional, and users can obtain a 3-month free subscription using the code PyCharm4HF.

AI products · Hugging Face Blog

Argilla 2.4 Release Introduces Zero-Code Construction of Fine-Tuning and Evaluation Datasets ↗

Argilla 2.4 introduces a feature that allows users to build fine-tuning and evaluation datasets on Hugging Face Hub without coding. It supports one-click import from any public Hub dataset, automatic generation of initial configurations, definition of questions, and real-time preview. Users can complete annotation, scoring, ranking, and other feedback collection tasks via the UI.