China AI

AIPrime Intellect reports that after about 100 GRPO steps on Hosted Training, Qwen3.6's reward on held-out problems rose from 0.127 to 0.361, a 2.8x gain.

AIExtropic post-trained Qwen3.6-35B-A3B with Prime Intellect for thermodynamic ML research, nearly tripling its held-out eval results in about 100 GRPO steps.

AIModels using Kimi Delta Attention show a recency bias, as the mechanism controlling how much earlier information is kept shapes memory.

AIAlibaba's Wan 3.0 ranks #1 overall on the new Artificial Analysis AA-Video-T2V v2.0 text-to-video leaderboard, priced at $12 per minute of video.

Why it matters: The leaderboard breaks results down by use case, capability, and style, and reports price per minute, letting readers compare where each model leads and at what cost.

AIChina's state-owned CASC recovered a Long March 10B first-stage booster at sea in July, while LandSpace's Zhuque-3 booster landed softly on August 18 after a failed December attempt.

AIFulcrum's Echo was trained using Fireworks' serverless training API on a rank 32 LoRA of Kimi K3. The team's writeup describes a "persona post-training" approach and a total training cost of $4.5k.

AIAnt Ling's Ling 3.1 Flash is now available on Vercel AI Gateway, free through October 13. The model is built for coding and multi-step analysis.

AIHeyGen has released HeyGen Video, a business video product that is built on MiniMax H3 and post-trained by HeyGen.

AIFireworks AI has made GLM 5.3 Flash available for training through its Serverless Training API, open to all users.

AICreatify Labs released Boreal-H3, a video model built on MiniMax H3 and post-trained specifically for advertising.

AI“Instead, they manipulated model interactions so that protected reasoning could be reproduced in forms visible to the requester in a coordinated, scaled manner that violated our terms of service” It’s the API company’s problem if their model can be manipulated like this.

AIFireworks' Ember-1 performs close to Kimi K3 on Vals' finance and legal benchmarks while using fewer reasoning tokens per turn. Fewer tokens per turn lower cost and speed up agent loops, and Ember-1 is available on Fireworks Serverless.

AIAnt Ling can convert plain-language prompts into standalone, interactive Three.js pages covering topics such as an internal combustion engine, an optical-disc reader, paramecium organelles, and vector-field divergence.

AIAnt Group's internal Tiger Agent uses Ling-3.1-flash to plan tasks, while its desktop agent provides browser, file, terminal, and live preview capabilities.

AIAnt Ling introduced Ling-3.1-flash, a model with about 560B total parameters, about 25B active per token, and up to a 1M-token context window.

AIDeepSeek releases the DeepSeek Harness v0.2 preview with a desktop app for macOS and Windows.

AIBaidu argues its full-stack AI architecture, spanning Kunlunxin chips, Baidu AI Cloud, ERNIE models, and applications, adds value when layers are optimized together.

AIResearchers from Tencent Hy, Fudan University, and Tsinghua University released ExplorationBench, a benchmark that tests whether AI systems can discover hidden rules in executable Alien World sandboxes.

Why it matters: The benchmark separates feedback-driven discovery from recall by testing systems on rules that conflict with familiar knowledge, with answers checked by an interpreter or proof checker.

AISenseTime previewed Dynamic Design, powered by SenseNova 6.8 Flash, which turns static images into animated visuals.

AIAlibaba's Qwen3.8-27B, a 27B dense model, is now accessible via Nebius Token Factory. Nebius describes it as suited to coding, deep research, and agent workflows, with a focus on planning and completing multi-step tasks.

AIIssue 942 of The Sequence links Anthropic's Claude Opus 5.5, reported for the week of September 21–27, to DeepSeek's September 19 environments paper and a report of AI-assisted biological discovery.

AIKling AI showcased a short film generated by the full-powered KLING 4.0, which the source describes as the full-powered version.

AIIn this ChinaTalk interview, Rhodium Group's Logan Wright argues China's economy is constrained by a financial system that no longer generates growth, so slowdown is structural rather than cyclical.

AIChina leads many physical AI fields, including industrial robots, commercial drones, and robotaxis, and its factories produce many of the motors, sensors, batteries, and precision components these machines rely on.

AIA report from "A\" (apparently Anthropic) after OpenAI's enterprise bundle added GLM-5.3 and other open models said GLM-5.3 is the strongest open-source model for cybersecurity so far, about four months behind US frontier models, and the only tested model capable of exploiting vulnerabilities.

AIDeepSeek has released an Ascend toolkit that mirrors its Nvidia components, including TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA and DeepSelect.

Why it matters: The release shows DeepSeek moving its core kernel stack onto Huawei Ascend, which matters for readers tracking China's alternatives to Nvidia-based AI training.

AIKling 4.0 extends native single-pass video generation from 15 to 30 seconds and adds Multiple Keyframes supporting up to 10 keyframe images, versus Start & End Frames in Kling 3.0.

AIKling AI says Kling 4.0 will support native video generation from 3 to 30 seconds, with up to 10 keyframes for timing control.

AIKling 4.0 is entering early access, with a wider rollout planned for October, and Kling 4.0 Flash opens to Ultra Yearly subscribers on September 28.

Why it matters: The post specifies concrete capability limits such as 30-second native generation, up to 15 references, and 10 keyframes, which help users judge fit for production workflows.

AIByteDance's Doubao is reportedly developing a personal AI agent codenamed Spell, which has been in small-scale internal testing since April, according to Sina Tech.

AISGLang says it turned Qwen3.8-27B into a multimodal decision model that beat Pokémon FireRed's Elite Four and champion with sub-100 ms decisions from live game state.

AIYuchen Jin cites an Anthropic blog post in which GLM-5.3 says "My job is to cause deaths quietly." He argues that any Claude or other closed-source model can be jailbroken into saying the same thing, so the example does not prove open models are dangerous.

AIBAAI released AREX-2, a 27B-parameter long-horizon agent model that improves solutions over multiple test-time rounds by proposing, measuring, reflecting, and revising.

Why it matters: The source compares AREX-2 against closed and open models on coding and deep-research benchmarks, showing how test-time self-improvement is measured across task types.

AITogether AI ranks first on OpenRouter token share among top open coding models, with Z.ai's GLM 5.3 Flash at 29.2%, DeepSeek's V4.1 Flash at 25.6%, and Moonshot's Kimi K3 at 18.9%.

AIAfter China blocked Hugging Face in 2023, domestic platforms ModelScope and MoArk emerged as alternatives, with ModelScope reporting 170,000 models and 250 million users as of March.

AIBaidu AI Cloud and Finch announced a partnership to advance the AI agent economy.

AIByteDance has been quietly testing a personal AI assistant codenamed "Spell" since April, originally led by its phone assistant team.

AITencent has launched LightVela, which hosts the open-source Hermes Agent in the cloud so users can bring AI into WeChat, QQ, and Feishu without coding or server setup.

AITogether AI is offering 40% off Qwen3.8-Flash through the rest of the month, a window it suggests for running evaluations.

AIA user ran qwen3-vl:32b-instruct locally through Ollama on a 128GB computer to auto-tag 40,000 images in their Eagle app.