China AI

AINathan Lambert says US open-weight models were behind Chinese models dating to February 2026 until today. He says today's release cut that gap in time by at least half, to roughly three months.

AIVercel introduces Token Pass, which gives $80 in GLM 5.3 and Kimi K3 tokens for $30 per month.

AINick Cammarata, formerly of OpenAI's interpretability team and now at Goodfire, says neural networks are not tangled knots and that polysemanticity is wrong.

AIKling AI says Kling 4.0 is now fully available through a refreshed creation interface at kling.ai, plus MCP and CLI access for video creation workflows.

AISword Health built Dawn 1 by reinforcement learning post-training GLM-5.3 on prime-rl, Prime Intellect's framework.

AIOpenRouter says teams in its private beta are running dedicated GLM, Kimi, and DeepSeek deployments, plus reserved OpenAI and Anthropic capacity. The post links to a blog announcement on how to try the private endpoints.

AIMustafa Suleyman says Microsoft matched DeepSeek-V4-Pro-Base's pre-trained score on held-out code using 12x less training compute. He attributes the result to compounding efficiency gains.

AINvidia has released Nemotron 3.5 Super VL, a 120-billion-parameter open-weight model aimed at agentic work.

AIHuawei says its Next-Generation (NG) WAN Solution has entered large-scale commercial use and is deployed by more than 60 carriers worldwide, announced at UBBF 2026.

AIStepFun's open-weights Step 5 Preview, a 600B sparse MoE model with 27B active per token, is now served at FP8 on Featherless.

AIBlack Forest Labs says FLUX 3 Image Klein scores higher than Qwen Image 2.1, the current open-weights leader, in text-to-image and multi-reference generation in its internal evals.

AISemiAnalysis highlights a GLM-5 paper claiming full-stack adaptation to seven domestic Chinese chip platforms, including Huawei Ascend, Moore Threads, Hygon, Cambricon, Kunlunxin, MetaX and Enflame.

AIKling AI says Kling 4.0 is now fully available. The company posted a sample "King Card" video, created with Kling 4.0, to show the model's output.

AIKling AI says Kling 4.0 is now fully available, accessible through a refreshed creation interface at kling.ai and through MCP and CLI for video creation workflows.

AIvLLM says DeepSeek-V4-Flash showed zero logprob difference at every step over 100 RL steps on 32 GB200 GPUs.

AIStepFun says its Step 5 Preview delivers strong inference performance on NVIDIA GB300 GPUs. NVIDIA says the 600B MoE model runs at nearly 8K tokens per second per GPU using NVFP4, and that it is StepFun's largest open-weights release.

AIStepFun has released Step 5 Preview, a sparse MoE model with 600B total parameters and 27B active per token, as open weights.

AINVIDIA AI congratulates StepFun on releasing Step 5 Preview as open weights, its largest open release.

AIBAAI says token-by-token autoregression in LLMs is strictly serial, so output frequency stays locked to single-token latency however much compute is added.

AIBAAI says a single stream can meet target SLOs on fewer GPUs when extreme minimum latency is not required.

AIBAAI says PhysisServe, evaluated across five ACWM and WAM backbones with full weights and no pruning, compression, or quantization, reaches 4.1 ms single-stream output latency for the 33B MiniMax H3 with one GPU per stage.

AIStepFun has released Step 5 Preview as open weights on Hugging Face and ModelScope, in BF16 and FP8 formats.

Why it matters: The post names the model's size, context window, and deployment paths, which helps readers gauge whether it fits their agent or serving workflow.

AIAlibaba's Tmall says its Double 11 shopping festival, which starts October 20, has drawn 1 million brands and 72 million products, both record highs.

AIAbdulrahman Alfozan, an OpenAI engineer, was arrested and charged after US prosecutors said he hid an encrypted recording device behind a conference room TV at OpenAI's San Francisco office in late September.

AIDeepSeek apologised and will issue users full refunds in credits for consumption between October 8 and October 14 after a nearly five-hour outage on Thursday.

AIAlibaba says a record one million global brands will take part in its Singles' Day shopping festival on Taobao and Tmall, with more than 70 million products expected to feature.

AIStepFun says BF16 and FP8 weights for Step 5 Preview are now available on Hugging Face and ModelScope, alongside a technical blog.

AIStepFun says its Step-5-preview model can be served through an OpenAI-compatible API with vLLM or SGLang, or run offline with Transformers, and it is adapted for multiple chip platforms and inference frameworks.

AIStepFun's technical blog describes how it co-designed the architecture and training of Step 5 Preview for longer, more complex agent tasks.

AIStepFun has updated Step Code to v0.1.4, its open-source agent harness built around Step 5 Preview.

AISGLang announces Day-0 support for StepFun's Step 5 Preview, a 600B-parameter sparse MoE model with 27B active per token and a 1M-token context window.

Why it matters: The post names the inference and training setup, including hybrid attention and Day-0 RL support, showing how the open-weight model is served and trained.

AISGLang (@sgl_project) shares a cookbook for running StepFun's Step-5-Preview model.

AISK Hynix has fallen 26%, Samsung 22% and Micron about 11% over five days as investors weigh CXMT's planned server memory product and a Maeil Business Newspaper column warning of oversupply in 2029.

AIAnt Ling thanks the vLLM project for making Ling-3.1-flash runnable on developers' own infrastructure from day one. vLLM announced day-0 support for the model, which has about 560B total parameters, about 25B active per token, and up to a 1M-token context.

AIAnt Ling's Ling-3.1-flash is available on DeepInfra from launch day, according to the post.

AIAnt Ling thanks Atomic Chat for running Ling-3.1-flash locally and sharing a LEGO Eiffel Tower demo. In Atomic Chat's post, Ling 3.1 Flash, released as open weights, beat GPT-6 Luna on the same Three.js prompt at 2.8x lower cost.

AIToken Rhythm says more than 300,000 developers have joined its platform and launched a Double 11 presale offer.

AIWannatech launched the NOVA21 dexterous hand on October 15 at the 2026 Hong Kong Autumn Electronics Fair, targeting robotics research and development teams.

AIAccording to people familiar with the matter, Alibaba's Qwen AI will be built into Apple Intelligence for users in China across iOS, iPadOS, macOS and visionOS. It would offer text and image understanding and content generation without app switching.

AISammy Azdoufal found that Energize Lab's Eiliko robot toy uploads users' conversations to servers in China, where he could read 13 million conversations and access a fleet of 68,000 devices.