China AI

AIDongxi NLP reports that DeepSeek has open-sourced DeepInfer, a high-performance, large-scale inference framework, under the Apache-2.0 license.

AIDeepSeek releases DeepInfer, its in-house inference framework, as open source on GitHub.

AIDeepSeek has released DeepInfer, its in-house inference framework, as open source on GitHub.

AIOpenBMB thanks @aiwithsally for demonstrating MiniCPM-V 4.7 in an automated report-generation workflow.

AIThe openbmb/tiny-processor-minicpm-ocr repository on Hugging Face is publicly listed, but users must accept its conditions and share their Hugging Face information before viewing its files.

AIThe Oct 20 brief from Testing Catalog lists ChatGPT Slides real-time collaboration for Pro, Business and Enterprise plans, and Meta's technical report on its 30B Apache 2.0 open-weight agentic model Muse Glimmer, which runs on a single consumer GPU. It also reports MiniMax M3.1 Flash launching at $0.15/$0.60 per 1M tokens with open weights due Oct 23, and Inception's Mercury 3 diffusion LLM running over 1,000 tokens per second with a 1M-token context window.

AINeolix, an autonomous delivery vehicle company, says it has closed a Series E round of nearly $400 million led by Boyu Capital and a national-level fund.

AISaidou Technology's first car, the AIVA ME7, appeared in a Ministry of Industry and Information Technology filing, with a 4,999mm length, 3,050mm wheelbase and CATL ternary lithium battery.

AIAlibaba's DAMO Academy introduced the XuanTie E908A, a RISC-V CPU IP core for automotive chips, at the RISC-V China Summit on October 19.

AISpacemiT and the XiangShan team introduced the V100 RISC-V server CPU alongside three complete servers at the RISC-V China Summit on October 19.

AIChina Telecom's XingChen AGI lab open-sourced Xing4.0-29B-Intent, a 29B model that classifies user intents and extracts their parameters in one pass, returning unified JSON. The team reports end-to-end inference latency under 150 milliseconds and a zero-shot cold-start Micro-F1 of 70.27, which it says beats GPT-4.1.

AIWuXi Stars Micro System Technologies introduced three interconnect chips for AI servers and clusters at its first product launch on October 15.

AIThe Department of Homeland Security ordered federal civilian agencies to remove DeepSeek's Chinese AI model and associated third-party services from their systems, citing national security concerns.

AIThe Department of Homeland Security made public an order barring federal civilian agencies from using DeepSeek products and services and requiring removal of DeepSeek-related technology from their IT systems.

AIPeking University and Qiyuan Robotics propose HIL-UMI, a method that collects error-correction data for embodied models without a robot arm.

AIAlibaba's Wan account says Wan3.0 Prime is first and Wan3.0 is third on OpenRouter's video leaderboard, New & Trending view. The account credits the placement to creators choosing Wan and asks users to share what they make with it.

AIThe weekly AINews roundup covers DeepSeek V4.1 Flash 1017, MiniMax M3.1 Flash, and Mercury 3, and notes that the first two are not yet in wide use.

AIChina Telecom's Xing general AI lab releases and open-sources Xing4.0-29B-Intent, an intent-understanding model, along with 2.75 million training corpus entries and model weights.

AIPhysis and vLLM introduce PhysisServe, a serving system that uses stage-wise pipelining, time-sharing, and joint multi-stream scheduling to improve both single-stream latency and multi-stream efficiency.

AIKling AI releases its Kling 4.0 video model, which generates single videos up to 30 seconds long.

AIAnt Ling thanks @AdinaYakup for sharing Ming-3.0 and highlighting its five open speech models. The post says the models are now available for developers to explore on Hugging Face, where Ant Group released its new speech and audio models.

AIPerch AI now offers MiniMax M3.1 Flash free of charge, alongside GLM 5.3 Flash, DeepSeek V4.1 Flash, and Beam 1.0.

AIWorkBuddy says it built a feature for users juggling Hong Kong and US positions across multiple screens.

AIAnt Ling thanks Atomic Chat for helping bring its Ming-3.0 speech models to local AI workflows.

AIAnt Ling thanks vLLM for day-0 vLLM-Omni support that makes Ming-3.0 easier to deploy. vLLM says Ant Ling open-sourced five speech models for TTS, audio understanding, and ASR, ranging from 0.5B to 126B total parameters with 6.5B active.

AIAnt Ling thanks the SGLang project for day-0 SGLang-Omni support and cookbooks for its Ming-3.0 models.

AIAnt Group's Ming-3.0-tts ranks first among open-weights models on Artificial Analysis's Controlled Voice leaderboard, with an Elo of 1,069, and ties Fish Audio's Drama 3 Preview for first on Pronunciation Robustness at 93.2%.

AIZhongke Tongliang showed its SmarCo AI workstation, SmarCo intelligent video encoding appliance and SmarCo AI acceleration card at the 6th RISC-V China Summit in Shenzhen on October 18.

AIMiniMax released M3.1 Flash, which it says costs one-third as much as M3 and runs about 3x faster.

AIVideoJev-4B, a Qwen3.5-4B-based model from MCG-NJU, scores candidate answers and time spans directly for video question answering and temporal grounding.

AIAnt Group's inclusionAI team has open-sourced Ming-3.0-tts under the MIT license, with a 0.5B dense model and a 7.9B-total, 1.3B-active MoE tiny model.

AIAgiBot says its Lingxi X2 humanoid has completed 46,686 interactions at the outpatient guidance desk of Peking University People's Hospital in Beijing as of September 30.

AITencent SSV is giving teachers across China free access to WorkBuddy's education edition, with 2,000 credits per month for 12 months and no regional or school-level restrictions.

AITencent's AI assistant Marvis is now listed in the Kylin software store for Kylin Desktop Operating System V10 SP1 on x86 and ARM, supporting multiple domestic CPUs.

AIAnt Group releases Ming-3.0-tts, an open-weights text-to-speech model under the MIT License, with a 0.5B dense version and a 7.9B mixture-of-experts variant named Ming-3.0-tts-tiny.

AIOllama says MiniMax M3.1 Flash is free on Ollama Cloud for the next three days, in collaboration with MiniMax.

AIThe Department of Homeland Security ordered all federal civilian agencies to remove DeepSeek products and services, including third-party services built on the app, and bars future purchases.

AIDeepSeek-V4.1-Flash-1017 (Max) scored 1670 on Arena's Code Arena WebDev leaderboard, making it the highest-ranked open-source model there, ahead of Kimi K3 (Max) and 12 points behind GPT-6 Sol (Max).

AIIndigo says the main barrier to agent interoperability is political rather than technical, pointing to walled-garden apps in China and US firms restricting agent access.

AISGLang announced a Low-Latency Mode built with NVIDIA and AMD that raises P90 per-user token speed on DeepSeek-V4.1-Flash.