China AI

AILi Auto says all new models launched from 2026 will use its self-developed batteries, which are already in the new L8, L6 and i8 and will reach the new MEGA, i9 and 2026 i6 in the fourth quarter.

AIKling AI's blog shows Kling 4.0 generating surreal bathroom scenes with a giant duck while keeping characters, facial expressions and camera movement natural.

AIRenting a Unitree G2 robot dog or G1 humanoid cost about $70 a day during this year's National Day holiday, down from $1,110 to $2,080 for a G1 in 2025, according to Jiemian News.

AINIO founder Li Bin told staff in an internal speech on September 8 that the company will launch a heavyweight new car next year.

AIZhijie Auto executive vice president Zhao Changjiang says the RX SUV, which opened pre-sales on August 20 at 299,800 to 399,800 yuan, is meant to become Zhijie's "Pura X," a new product anchor above the R7.

AINIO founder Li Bin says the brand's average transaction price reached 406,000 yuan in Q2 2026 and 434,600 yuan in July, about 180,000 yuan above Audi, 70,000–80,000 yuan above BMW, and several tens of thousands above Mercedes-Benz.

AIChina's AI drama and animation market exceeded RMB 22 billion ($3.29 billion) in the first five months of 2026, according to DataEye's H1 2026 report, with 221,900 new AI short dramas released from January to June.

AIChina's State Council Information Office says it will roll out a national plan to create jobs and build workforce AI skills in response to AI development.

AIThe WorldArena 2.0 Challenge, run by a Tsinghua University-led team at IROS 2026, ranked world models across three tracks.

AIXiaomi's MiMo-V2.6 technical report argues that scaling reinforcement learning, not more pretraining data, is the main lever for frontier capability, along batch size, environment diversity, and grader strength.

Why it matters: The piece walks through the report's three-way RL scaling method, batch size, environments, and grading, with concrete failure modes and stabilization fixes useful to agentic RL practitioners.

AIAlibaba's AI coding agent Qoder has added a Fast mode with a first response time of about three seconds and a credit rate of 0.8x, live in Qoder desktop app version 0.4.4 and later.

AIStriding AI unveiled H1 and C1 robots for round-the-clock convenience stores at APRCE 2026 on October 8, with commercial service targeted for 2027.

AIModelScope announces VeriLoop E2, a 27B open-weight model built on Qwen3.8-27B with a 262K native context window.

Why it matters: The post names the open-weight base model, the verifier-governed training loop, and the benchmark results, so readers can check the claimed gains against comparison models.

AIMicrosoft releases Microsoft-Decision-1, a decision-scoring model post-trained from Alibaba's Qwen3.5-9B that returns calibrated probabilities for fixed answer options instead of generated text.

AIStepFun's flagship Step 5 Preview ranks first on OpenRouter's Trending chart and second on its daily chart by token volume, and will be open-sourced on October 15, the company says.

AIAlibaba chairman Joe Tsai said open-source AI is the only way for Europe to achieve technological independence, citing China's open-source push as a model.

AIMiniMax says its M3 model now runs on NVIDIA Vera Rubin NVL72 systems using vLLM, with support from vLLM, Inferact, NVIDIA, and Red Hat.

AISeres says it sold 29,271 new energy vehicles in September, up 20.7% month on month. The new-generation Aito M9 series has delivered more than 40,000 units in 16 weeks, and the M9 Ultimate extended edition began deliveries on September 12.

AIA Tsinghua, Oxford, and Stanford paper finds that LLMs still write out reasoning with thinking turned off, especially on open-ended questions.

AIXiaomi's MiMo-V2.6 paper says agents now build tasks, audit tests, grade answers and detect cheating during RL training, with humans setting the budget and rules.

AIChina's average daily token calls passed 140 trillion as of March 2026, according to Samsung's Kevin Yoon, who spoke at GMIF 2026.

AIQuite a delight to see such a wonderful video on Saturday morning!

AITsinghua researchers present TokenRouter, a serving system for token-level routing between small and large models that raises throughput 2.01 to 64.15 times over the stronger existing setup across five routing methods.

AILenovo's TianxiCode coding agent, running DeepSeek-V4.1-Flash, resolved 71% of tasks on the SWE-bench-Live Lite leaderboard and passed official verification, according to a Lenovo statement.

AIPandaily relays an Internet Jianghu commentary that sums up four Chinese phone AI strategies: Huawei goes deep, Alibaba goes broad, ByteDance goes focused and Tencent goes clever.

AIShengShu Technology has released Vidu Q4 Preview, a first preview of its next flagship video generation model, supporting up to 15 reference images, up to three reference audio clips, and output up to 4K with 10-bit color depth.

AISGLang has been optimized for NVIDIA Vera Rubin to accelerate Kimi K3 inference, working with NVIDIA on attention, MoE, and speculative verification kernels on early-access hardware.

AIInferact says vLLM now supports NVIDIA Vera Rubin, with early results showing more than 7.8x the throughput of GB200 on MiniMax M3 at matched interactivity on AgentX. The team says it integrated a Rubin-optimized MSA prefill kernel and locality-aware optimizations, and describes the results as early.

AIInferact, working with NVIDIA, led performance tuning and integrated the MiniMax Sparse Attention kernel.

AIvLLM now supports NVIDIA Vera Rubin, according to the vLLM project's X account.

AIvLLM says it delivered 7.8x or more the throughput of GB200 on SemiAnalysis AgentX, serving MiniMax M3 on Vera Rubin NVL72 at matched interactivity.

AIvLLM says users can try it on Rubin today with the vllm/vllm-openai:cu134-nightly image, which runs DeepSeek, Kimi, GLM, and MiniMax models. The post links to a blog preview and says further improvements for Vera Rubin are planned.

AITinker has cut prices up to 70% on long-context prefill and sampling, which now cost the same as short context.

AISGLang says it worked with NVIDIA to optimize attention, MoE, and speculative verification kernels for Kimi K3 inference on early-access Rubin hardware.

AIHiDream-O1-Video-1.0 ranks #6 in Artificial Analysis's Image to Video with Audio leaderboard, just behind Dreamina Seedance 2.0 720p.

AIA GitHub post announces that kernel kits for DeepSeek's TP=2 and TP=4 recipes are now live. The post links to a repository named deepseek-v4.1-tensorfold-tp2-2xgb10 and gives no further details on performance or features.

AITinker says it made efficiency improvements to support scaling long-context reinforcement learning and is passing them on through price cuts of up to 70%.

AITinker says prefill for 128k and 256k context no longer costs extra, an effective discount of over 2x for long-context models including Kimi K2.6, gpt-oss-120b, and Inkling.

AITinker adds GLM-5.3-Flash and DeepSeek-v4.1-Flash, both of which natively accept image inputs and use efficient attention architecture.

AIStepFun's Step 5 Preview, a 600B-total, 27B-active MoE model with 1M context and vision, is free to try on Nous Portal for one week. Nous Research says it scored 33.89 on the Hermes Index, the same score as GPT-6 Luna.