Foundation Models
LLMs & AI modelsChina's foundation models — Doubao, DeepSeek, Qwen, Kimi, GLM, Ernie, Hunyuan and more: releases, capabilities, open-source moves and benchmarks.
Qwen Powers Apple Intelligence in China via Siri, Writing Tools and Visual Intelligence
Apple Intelligence has opened for testing in China through the latest iOS developer beta, with Alibaba's Qwen handling cloud-side complex generation and text-image understanding across Siri, Writing Tools and Visual Intelligence. The integration is opt-in, and Apple plans a general rollout to users in China in December.
Tencent Hunyuan Releases Hyra 2.0, But Only Protein Models Come With Code
Tencent's Hunyuan team released Hyra 2.0 on November 2, a system for autonomous AI research that lists seven results spanning proteins, mathematics, quantum codes, a robotic hand and RISC-V chip optimization. Only the two protein models ship with weights, code and a technical report; the rest rely on self-evaluation, simulation or remain closed.
LongCat-2.5: Meituan Claims Frontier-Adjacent Performance from Domestic Compute Only
Meituan has released LongCat-2.5, a natively multimodal 1.6-trillion-parameter model it says was trained end to end on domestic computing power, with reported gains of 86% in per-chip training throughput and about 40% in inference throughput. The company positions the model between two Claude Opus generations and is channeling it into merchant-facing tools rather than a public token business.
Half the World's Tokens Now Run on Chinese Models — and Half of China's Go to Short Dramas
OpenRouter data cited in the article puts Chinese models at more than half of global token consumption, while a Chinese Online filing attributes 55% of domestic token use to AI short dramas and video generation. The piece traces how platform subsidies from Hongguo to WeChat's Green Bubble turned short-drama production into a compute-cost business, in contrast to the US, where coding dominates token demand.
Token Prices Rebound to Pre-Price-War Levels While Call Volume Quadruples, CAICT Report Says
A new CAICT monitoring report says China's average daily large-model token consumption reached about 500 trillion in June 2026, nearly four times the December 2025 level, while the normalized average price rose from 6.8 yuan to 13.46 yuan. Domestic throughput doubled to 78.23 TPS but still lags typical overseas models at 112.82 TPS.
Moonshot AI Releases Kimi K3.1 Pro, Betting on Post-Training Compute Over Bigger Models
Moonshot AI released Kimi K3.1 Pro on October 24, a post-trained version of Kimi K3 with unchanged 2.8T total parameters that the company says ranks first among open-source models on the Artificial Analysis index. The upgrade targets coding agents, 3D model generation, and finance and legal workflows, with full weights due by November 4.
Iflytek Pushes 60% of AI Tasks On-Device With New Edge-Cloud Split
At its 1024 Developer Festival in Beijing, Iflytek laid out an edge-cloud allocation scheme that runs 60% of routine tasks on-device, 30% collaboratively and 10% in the cloud. It also released the Spark X2.6 model, a 293B-A30B MoE system with a 1M context window, plus an open-source 4B on-device version.
DeepSeek Details V4.1 Flash Inference Stack: Split Prefill/Decode and Week-Long KVCache
DeepSeek has open-sourced DeepInfer, the inference framework behind its V4.1 Flash online service, along with a token encoding/decoding component called deepseek-recipe. The disclosure details a disaggregated architecture with eight-card Prefill servers, 64- to 128-card Decode supernodes, and a cluster-level KVCache layer, though KVCache management and load-balancing code remain unreleased.
MemT 1.0: MemTensor's Memory-Native Models Run a Memory Path Alongside Attention
Shanghai-based MemTensor has open-sourced MemT 1.0, a pair of memory-native foundation models (288B Flash and 36B Nano) that embed persistent, updatable memory inside the Transformer alongside attention. The team reports top rankings among parametric memory methods and large inference speedups on long histories, though scores still trail its own text-based MemOS system.
MiniMax M3.1 Flash Cuts Total Usage Cost 70%, Targets Agent Workloads
MiniMax released M3.1 Flash on October 19, confirming that the anonymous model Space Bunny was an early version tested across more than 200 trillion tokens. The company claims nearly 3x faster inference and a 70% cut in total usage cost, with API prices down to 1 yuan per million input tokens.
Against the Price War: StepFun Raises Step 5 Preview API Pricing Ahead of Weight Release
StepFun doubled the cache-hit input price for its Step 5 Preview API to 0.70 yuan per million tokens from 00:30 Beijing time on October 14, 2026, leaving cache-miss input and output unchanged, one day before open-sourcing the model's weights. The move runs against three years of Chinese large-model API price cuts and, by analysts' reading, reflects tight compute supply and an attempt to set a price floor against customer self-hosting.
SenseTime Upgrades Yunyu Xingkong Model as Shanghai Deploys First Quantum City Government AI Agents
SenseTime's SenseCore and the Shanghai Municipal Bureau of Planning and Natural Resources have upgraded the Yunyu Xingkong large model, billed as China's first industry foundation model for planning and natural resources, strengthening spatial reasoning and task execution. The bureau also released its first batch of Quantum City government AI agents covering flood-control dispatch, spatial analysis and planning updates.
Zhipu's Tang Jie Lays Out GLM's Three Scaling Directions on the Road to AGI
Speaking at the National University of Singapore, Zhipu founder and Tsinghua professor Tang Jie outlined three scaling directions for GLM — pre-training, post-training, and reasoning plus agents — and named recursive self-improvement and robot action as the two capability gaps still to close. He also highlighted long-term memory, rather than longer context windows, as the key to continual learning in the agent era.
Qwen-Audio-3.1 Adds Five Models Spanning Audio Understanding, Generation and Interaction
Alibaba's Qwen team released the Qwen-Audio-3.1 series on September 23, adding five models covering speech recognition, synthesis, real-time interaction and audio creation. Prices were cut across the board, with ASR down as much as 95%.
Tencent Hunyuan Launches Hy Image3.5 Preview With Multi-Turn Creation
Tencent's Hunyuan team has launched the Hy Image3.5 preview model for professional-grade image generation and editing, supporting text-to-image, image-to-image, and multi-turn conversational creation. The model is already live across Yuanbao and several partner apps, with API access on Tencent Cloud's TokenHub.
Xiaomi Open-Sources MiMo-V2.6, Scaling Reinforcement Learning for Model Self-Improvement
Xiaomi has released and open-sourced its MiMo-V2.6 series, two full-modality models built around large-scale reinforcement learning to explore recursive self-improvement. The company says MiMo-V2.6-Pro ranks among today's stronger open-weight models while keeping prices unchanged.
Qwen Open-Sources Qwen-Image-2.1, a 7B Unified Image Generation and Editing Model
Qwen has open-sourced Qwen-Image-2.1, a 7B-parameter model that unifies text-to-image generation and image editing, with native RGBA transparent output, up to 10 reference images, and mask-based local editing. The release also adds KV cache reuse for inference efficiency and targets portrait and product consistency.
Shengshu Technology Launches Vidu S2 with Real-Time Video Generation, Editing and Spatial Video
Shengshu Technology has released Vidu S2, a real-time video model that comes in two variants: S2-Avatar for interaction with digital characters, and S2-Editing for editing video streams. The new model supports real-time 720P output, lets users inject reference images mid-generation, and can adjust clothing, subjects, backgrounds and visual style in a video on the fly. The team also demonstrated spatial video generation and editing aimed at VR headsets.
Doubao Mobile Assistant Adds Screen-Based Tasks, Personal Memory and Cross-App Actions
ByteDance has launched the consumer edition of Doubao Mobile Assistant with expanded screen interaction, on-device information retrieval, personal memory and cross-app task execution. The assistant will debut on nubia's NaviX Ultra, while ByteDance says reliability, GUI agents and cross-app integration still need improvement.
Xiaohongshu's AllSpark Releases Iris: Leading Performance Among Open-Source Search Agents in Its Class
Xiaohongshu's AllSpark team open-sourced Iris Search Agent weights and evaluation code, with Iris-mini and Iris-pro topping BrowseComp, BrowseComp-ZH, DeepSearchQA and HLE benchmarks in their respective parameter classes.
Un, the First Large-Scale Generative Model to Use Physics as a Computational Primitive
Over the past decade-plus, GPU-centric digital computing has dominated AI. Bigger clusters, higher bandwidth, more powerful GPUs, and denser data centers have seemed to be the mainstream path toward the next generation of AI.
DeepSeek and Kimi Diverge as Funding Rises
Liang Wenfeng and Yang Zhilin Start Answering Different Questions
ByteDance Lights Up Its Own ChatGPT Moment During Chinese New Year
The real AI offensive is a hard decision to turn itself into a technology company.
Doubao’s 500-yuan-a-month Pro Plan Is Officially Here. Is It Expensive?
Can an Agent-Driven Office Workflow Really Replace Employees?
Huawei’s SpaceMind Tops a Leading Spatial Intelligence Benchmark: A Pure RGB Vision-Language Model Scores 70.6, Setting a New Record on Fei-Fei Li’s Leaderboard
Large models can now hold fluent conversations and recognize objects in images, but a more fundamental question remains unresolved: do they actually “understand” the three-dimensional world we live in?
Huawei Open-Sources 7B Multimodal Model With Strong Visual Grounding and OCR, Bringing a New Ascend Edge “Sweet Spot
Models in the 7B parameter class have long been a favorite for edge deployment and individual developers.
Kimi K3 Sparks Heavy Demand as GPUs Struggle to Keep Up
Kimi K3 is drawing strong attention, with demand apparently outpacing available GPU capacity. The headline points to pressure on compute resources around the model, but the body provided is empty and details cannot be verified.