China AI

AIInferact and Google submitted their first preliminary TPUv7 results for the 2.8T-parameter Kimi K3 model.

AIKevin Weil, OpenAI's account owner, posts only an eyes emoji in reply to a post by @m_goes_distance.

AIArena says DeepSeek-V4.1-Flash-20261017 (Max) is now the #1 open model in Code Arena WebDev, ahead of Kimi K3 (Max).

AIMerve Noyan says she could not download llama.cpp directly, so she built a version using wllama by @ngxson to run a 4-bit quantized Qwen3.5-0.8B on the Rabbit R1.

AIMerve Noyan of Hugging Face says the economic value created by open-source AI is unmatched.

AIThomas Wolf, Hugging Face's CEO, estimates that Meta, NVIDIA, Alibaba, DeepSeek, Moonshot, Mistral and about 30 other labs have shared close to $1B of training compute on the Hub.

AISemiAnalysis analysts discuss GLM 5.3, an open-weight model from Z AI, in a podcast episode, noting its architecture is unchanged from 5.2 and its agentic benchmark gains of over 2% come from post-training.

AIDeepSeek releases V4.1-Flash-1017 with MIT open weights, according to the October 17 edition of AI Compass. The same roundup reports that OpenAI launches ChatGPT Slides and that Claude traces the origins of the NIST ECDSA curve seeds.

AIRokid opened an unstaffed flagship smart-glasses store in Hong Kong's Harbour City, where customers test devices, pay on screens and collect products from a robotic system without any sales staff.

AISUPCON has adopted Alibaba's QwenWork enterprise AI agent, starting with its marketing chain from lead collection to payment collection, under a strategic cooperation agreement signed October 16.

AIGlenfly Tech says its first Arm-based system-on-chip, the GF800, has finished chip debugging and is sampling to end customers.

AIThe 2026 Bay Area Semiconductor Industry Ecosystem Expo (WESEMiBAY) closed on October 16 at the Shenzhen Convention and Exhibition Center (Futian), with over 156,700 professional visitors across nearly 80,000 square meters.

AIDeepSeek updated V4.1 Flash to V4.1-Flash-1017, and its web, app and API models now run the new version.

AIDesign Arena says DeepSeek-V4.1-Flash-1017 is now available on its platform.

AIWorkBuddy says DeepSeek-V4.1-Flash-1017 is free through October 31 on its international version, while the domestic version has no free promotion yet.

AIKling AI's X account announces Kling 4.0 with the hashtags #Kling4 and #KlingAI, without listing features, specs, or a release date. The quoted post from idea.ai says Kling 4.0 will come soon to its platform.

AIPeng Cheng Laboratory has begun trial operation of 10,000P of domestically built AI computing on Peng Cheng Cloud Brain III in Shenzhen, announced at the 2026 World Computing Network Conference.

AIDeepSeek officially released V4.1-Flash-1017, a minor update to V4.1 Flash, with weights on Hugging Face and ModelScope under the MIT license.

AITsinghua's THUNLP lab, ModelBest, OpenBMB and AI9Stars released PilotDeck 2.0, an open-source agent operating system with isolated project workspaces and white-box memory.

AIWeChat for HarmonyOS version 8.0.22 now works with Huawei's voice assistant Celia through agent-to-agent (A2A) integration, letting users send WeChat messages and start voice or video calls to individual friends by voice.

AIvLLM reports Kimi K3 decode reaching 842 tokens/s at batch 1 on NVIDIA GB300 NVL72 using DSpark.

Why it matters: The post gives concrete throughput, interactivity and configuration results, showing how a unified decode kernel changes serving performance for a large model.

AINVIDIA CentML's Benjamin wrote decode-specialized kernels in Cute DSL for an upstream vLLM optimization, reaching 423 tok/s per user on the 2.8T-parameter Kimi K3 model.

AIDongxi NLP posts that DeepSeek-V4.1-Flash-1017 is released, dated 2026-10-17. The post's second line, (roughly "Don't be too good to me, I hate the gap"), gives no further detail.

AIDeepSeek announces an upgrade to its V4.1 Flash model, with DeepSeek-V4.1-Flash-1017 now available. The source gives no further details on benchmarks, pricing, or availability terms.

AIZ.ai's GLM 5.3 was built entirely by post-training its predecessor, GLM 5.2, without pretraining from scratch.

AIDubai's largest-ever delegation of more than 100 government and business leaders attended the Dubai Business Forum in China, where 10 cooperation memoranda were signed.

AITsinghua University's THUNLP lab, with partners including ModelBest and OpenBMB, has released PilotDeck 2.0, an open-source agent operating system.

AIZhipu has released GLM-Legal, an open-weight legal model that handles contract review, evidence analysis, and cross-border compliance tasks.

AIDeepSeek has released DeepSeek-V4.1-Flash-1017, which the company says improves benchmark scores while shortening task runtime and output length.

AILM Studio says DeepSeek-V4.1-Flash-1017 is now available in LM Studio Bionic, with a 1M context window, image input, and US-hosted processing under full zero data retention.

AIOllama says DeepSeek-V4.1-Flash-1017 is available on Ollama Cloud, hosted in the US with zero data retention.

AIZevio says DeepSeek V4.1 Flash 1017 is now available on its platform at 45% lower inference cost.

AIElvis Saravia says the performance and cost improvements are arriving quickly across AI providers and matter to builders.

AIThe author reports that DeepSeek released DeepSeek-V4.1-Flash-1017, a 763B-parameter model with open weights under the MIT license.

AIYuchen Jin says DeepSeek-V4.1-Flash-1017 scores 77 on DeepSWE v1.1, matching Opus 5-level performance at 83x lower cost. He adds that Databricks will launch the model.

AIGAC Aion's Ray7 coupe began pre-orders on October 17 at 128,800 yuan and up.

AIBolt is adding GLM 5.3 Flash to every account's model picker this weekend at no cost, with no plan change or code required. The post is promoted as part of Bolt Builders Week, which offers a $10k prize for apps submitted by October 20.

AIModelScope announces DeepSeek-V4.1-Flash-1017, which it says delivers stronger coding and agent performance with faster task completion and shorter outputs.

AITencent researchers present Terminal-RSI, which starts from Qwen3.8-27B and alternates two phases over five stages.

AIThe Guardian column argues Britain should reject cheap Chinese AI from developers such as DeepSeek and Xiaomi, citing concerns over state censorship and data sharing.