1. DeepSeek Open-Sources DeepInfer Inference Framework Powering DeepSeek-V4.1-Flash

DeepSeek has open-sourced DeepInfer, its in-house inference framework used to power DeepSeek-V4.1-Flash in production. The framework is designed for high throughput and low latency at scale. DeepSeek says releasing its production stack aims to help the community lower compute costs and reduce energy use. The move extends DeepSeek’s pattern of open-sourcing core infrastructure alongside its models.
Read more: https://mp.weixin.qq.com/s/5MEu2eeqNfBWXlkCKTd6sg

Video Credit: NotebookLM

2. Unitree unveils H2-X ‘Kung Fu Legend Guardian’ security robot with up to 8-hour endurance

Unitree Robotics introduced the H2-X, a humanoid robot marketed as the ‘Kung Fu Legend Guardian,’ designed for guarding and service tasks. The company states the robot offers a maximum endurance of approximately 8 hours. Unitree also urged users to operate the robot safely and in a friendly manner. The announcement was shared on Unitree’s official account alongside a demonstration video.
Read more: https://x.com/i/web/status/2112819774573179175

Video Credit: @UnitreeRobotics on X

3. Tencent’s Qoder launches Qoder Lite, a $10/month AI coding plan with multi-model routing

Tencent’s Qoder AI IDE launched Qoder Lite, a new $10 per month subscription tier offering 1,000 credits. The plan routes everyday tasks to Flash models at a reduced 0.1x credit cost, while more demanding tasks are handled by Qwen3.8-Max, DeepSeek-V4-Pro, or Kimi-K3. As a launch promotion, first-time subscribers receive an additional 1,000 bonus credits, doubling their first month’s allowance to 2,000. The tiered model routing aims to balance cost efficiency with capability for complex coding tasks.
Read more: https://mp.weixin.qq.com/s/rHSzJcyhH0GkTDcs4Wi5wg

Video Credit: NotebookLM

That’s all for today’s China AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.

The post China AI Native Industry Insights – 20261022 – DeepSeek | Unitree | Tencent | more appeared first on AI Native Foundation.

1. DeepSeek Releases V4.1-Flash-1017 with Faster, Cheaper Performance

DeepSeek announced the release of DeepSeek-V4.1-Flash-1017, an updated model now available on its website, app, and API. The company says the new version delivers stronger benchmark performance, faster task completion, and shorter outputs compared to its predecessor, representing a Pareto improvement in both performance and cost. The model is optimized for use with DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. DeepSeek recommends upgrading to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model.
Read more: https://mp.weixin.qq.com/s/X-5kbBkedQ5Oct9byUunag

Video Credit: The original article

2. DeepSeek Opens Search and Fetch APIs to All Developers

DeepSeek has opened its Search and Fetch APIs to all developers. The APIs are powered by DeepSeek’s in-house search engine, which handles hundreds of millions of requests daily worldwide. Each endpoint is priced at 0.75 US dollars per 1,000 calls. Calls made through server-side tools in DeepSeek Harness, Claude Code, and Codex remain free for DeepSeek API users until billing begins on October 31, 2026 at 16:00 UTC.
Read more: https://api-docs.deepseek.com/guides/search/

Video Credit: NotebookLM

3. Alibaba’s Ant Group Open-Sources Five Ming-3.0 Speech AI Models

Ant Group released five open-source Ming-3.0 models spanning three speech-related families: text-to-speech, audio understanding, and automatic speech recognition. The TTS family includes a 0.5B model for voice design and paralinguistic control and a 7.9B (1.3B active) model for long-form speech. The audio family includes a 126B (6.5B active) model for understanding and tool use and a 9.3B (2.7B active) model for efficient dialogue. The ASR family adds an 8.6B (2B active) model supporting dialects and contextual recognition.
Read more: https://mp.weixin.qq.com/s/UZ30ww14wXhgOJ4jqnCnKA

Video Credit: The original article

4. OpenBMB’s MiniCPM-V 4.7 ranks third on Hugging Face trending list

OpenBMB announced that its MiniCPM-V 4.7 multimodal model reached third place on Hugging Face’s trending list. The release includes a 35B-A3B mixture-of-experts model that runs smoothly on iOS and Android devices using about 3GB of memory. A smaller 1B parameter variant delivers strong performance on OCR, math, and video understanding tasks. The model and code are available on GitHub and Hugging Face.
Read more: https://github.com/OpenBMB/MiniCPM-V

Video Credit: NotebookLM

That’s all for today’s China AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.

The post China AI Native Industry Insights – 20261020 – DeepSeek | Alibaba | OpenBMB | more appeared first on AI Native Foundation.

1. Alibaba’s AntLingAGI open-sources Ling-3.1-flash with 560B parameters and 1M context window

AntLingAGI, part of Alibaba’s Ant Group AI efforts, has open-sourced Ling-3.1-flash, a large language model with roughly 560 billion total parameters and about 25 billion active parameters per token. The model supports a context window of up to 1 million tokens. It is designed for practical use cases, able to turn a rough brief into a webpage, mobile UI, or interactive prototype for teams to review. A demo was shared alongside the announcement.
Read more: https://huggingface.co/inclusionAI/Ling-3.1-flash

Video Credit: @AntLingAGI on X

2. OpenBMB and Tsinghua THUNLP Release Open-Source AI Agent Harness PilotDeck 2.0

OpenBMB, Tsinghua THUNLP, ModelBest, and AI9Stars released PilotDeck 2.0, an open-source, model-agnostic AI agent operating system that executes natural language instructions with transparent data flows. The update adds end-to-end office document handling for Word, Excel, and PowerPoint, more reliable execution of long-running tasks stress-tested across 2,000 interaction messages, and a one-click desktop installer. On the WorkBuddy Bench benchmark, PilotDeck scored 79.1 using the Qwen3.6-35B-A3B model, outperforming WorkBuddy’s 74.7 and Claude Code’s 77.3. The project remains fully open-source, letting users choose their own underlying models.
Read more: https://github.com/OpenBMB/PilotDeck

Video Credit: @OpenBMB on X

3. Qoder releases custom Agent and Agent Team features for individual users

Qoder has opened its Agent and Agent Team features to individual users, allowing them to configure custom AI agents with specific responsibilities, work requirements, and available tools. Users can create specialized agents for tasks like competitive research or code review by combining skills, MCP data sources, and plugins. For complex tasks requiring multiple roles, users can form agent teams led by a Leader agent that coordinates work across members, each operating in separate conversation contexts to manage complexity and reduce context window burden.
Read more: https://mp.weixin.qq.com/s?__biz=MzE5ODI2NzI5Nw==&mid=2247493374&idx=1&sn=ab854f04301e1d182cf0708d0a191f91&chksm=97c9186d231cbe026c870043a7ad34188d06d3a9f5625f799fea1fbe4f8177e21e7174629924&scene=0&xtrack=1#rd

Video Credit: The original article

That’s all for today’s China AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.

The post China AI Native Industry Insights – 20261015 – Alibaba | OpenBMB | Qoder | more appeared first on AI Native Foundation.

1. OpenBMB Releases MiniCPM-V 4.7, a 35B MoE Multimodal Model That Runs on Mobile Phones

OpenBMB has released MiniCPM-V 4.7, a 35B-parameter MoE multimodal model that can run on mobile phones with as little as 3GB of memory. The model uses a technique called Lookahead Expert Scheduling (LES) to enable efficient MoE inference on mobile devices, and a lightweight 1B variant is also available for resource-constrained hardware. It supports tasks such as reading real-world signs, parsing complex documents, and reasoning over visual data on both iOS and Android. A companion method, Self-Replay Fine-Tuning (SRFT), lets developers customize the model for new tasks while preserving general performance.
Read more: https://github.com/OpenBMB/MiniCPM-V

Video Credit: @OpenBMB on X

2. Alibaba Releases Qwen-Image-2.1-Turbo with Open Weights for 8-Step Image Generation

Alibaba’s Qwen team released Qwen-Image-2.1-Turbo, an accelerated checkpoint of the 7B-parameter Qwen-Image-2.1 visual generation model that creates and edits images in just 8 denoising steps. The model retains support for 2K text-to-image generation and natural-language editing despite the reduced step count. Open weights are available on Hugging Face and ModelScope, and the model can be used directly via the Diffusers library. Alibaba also launched official Pro and Turbo API endpoints for hosted access.
Read more: https://modelscope.cn/models/Qwen/Qwen-Image-2.1-Turbo

Video Credit: NotebookLM

3. Z.ai Launches GLM-Legal, a Legal-Domain AI Model Built on GLM-5.3-Flash

Z.ai introduced GLM-Legal, a specialized legal AI model built on its GLM-5.3-Flash foundation model. The company says GLM-Legal matches frontier models on mainstream legal benchmarks. The model is designed for real-world legal workflows including contract review, evidence analysis, and compliance checks. Z.ai published a technical blog post detailing the model and opened API access applications alongside a demo tool called AutoClaw.
Read more: https://z.ai/blog/glm-legal

Video Credit: The original article

That’s all for today’s China AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.

The post China AI Native Industry Insights – 20261014 – OpenBMB | Alibaba | Z.ai | more appeared first on AI Native Foundation.

1. Huawei Unveils OceanStor M900 Context Memory Storage for AI Data Centers

Huawei introduced the OceanStor M900 Context Memory Storage system, designed to address the growing demands of expanding AI context windows and large-scale inference workloads. The product offers massive shared storage capacity alongside high-performance data access. Huawei states the system is optimized to reduce token costs for hyperscale AI data centers. The release targets infrastructure providers scaling AI inference operations.
Read more: https://x.com/Huawei/status/2108496892086890587

Video Credit: @Huawei on X

2. Qoder AI IDE launches Fast mode, cutting response time and credit cost

Qoder, Tencent’s AI coding IDE, released a new performance mode called Fast. The company says it delivers the same output quality with response times around 3 seconds. Credit consumption for the mode drops from 1.1x to 0.8x compared to the previous rate. The feature is now live for users.
Read more: https://x.com/qoder_ai_ide/status/2108806953141162235

Video Credit: @qoder_ai_ide on X

3. CapCut Launches Timeline Editor in Video Studio on Web

ByteDance’s CapCut has launched a Timeline Editor within CapCut Video Studio on the web. The new tool combines AI video generation with manual editing in a single canvas, letting users trim and combine clips, refine prompts and regenerate content, add music, and make further adjustments in one workflow. Projects started on the web can be continued seamlessly on CapCut Desktop. The update is available now for users to try.
Read more: https://x.com/capcutapp/status/2108821764486914201

Video Credit: @capcutapp on X

That’s all for today’s China AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.

The post China AI Native Industry Insights – 20261012 – Huawei | Tencent | ByteDance | more appeared first on AI Native Foundation.

1. Vidu Releases Q4 Preview, a Next-Generation Flagship AI Video Model

Vidu AI launched Vidu Q4 Preview, described as its next-generation flagship AI video generation model. The company highlights expressive character performance, cinematic camera movement, and high-impact visual effects as key improvements. Vidu Q4 Preview is now live and priced from 0.014 dollars per second of generated video. The release aims to make flagship-level AI video generation more accessible through lower-cost pricing.
Read more: https://x.com/ViduAI_official/status/2107856382263505219

Video Credit: @ViduAI_official on X

2. Tencent releases ima, an AI knowledge workspace powered by Hunyuan and DeepSeek models for writing and knowledge management

Tencent has launched ima, an AI-powered knowledge workspace that integrates search, reading, and writing capabilities. The platform is powered by Tencent Hunyuan and DeepSeek dual AI engines and features a copilot function that learns user habits and preferences over time. The tool is available on Windows, Mac desktop platforms and WeChat mini-program, with cross-platform data synchronization. New users receive 500 computational credits upon registration plus 100 daily credits for regular use.
Read more: https://mp.weixin.qq.com/s?__biz=Mzk0MDg0MTkzOA==&mid=2247506267&idx=1&sn=a9b2d72305b83670c01b33a8f3fd5f3c&chksm=c32f81af78f15e16b67b3e3ca60e52333a2c457025a813402e2c7b119f5d0cc53604352fd90a&scene=0&xtrack=1#rd

Video Credit: NotebookLM

3. Kuaishou’s Kling AI releases version 4.0 with 30-second native video generation and enhanced stability for complex camera movements

Kuaishou’s Kling AI announced the beta release of version 4.0, scheduled for official launch in October 2026. The new model introduces up to 30-second native video generation and improves stability in complex camera movements and large dynamic scenes. It enhances understanding of long prompts and continuous narratives while maintaining consistency of characters, props, and scenes across shots. Additional capabilities include subject and scene asset invocation, sound continuation, and motion transfer. An advertising creative director with 28 years of experience tested the model and confirmed its improved performance in high-speed camera work, action sequences, cross-scene consistency, and character performance with emotional speech.
Read more: https://mp.weixin.qq.com/s?__biz=MzU5NTkwNDU2OA==&mid=2247496912&idx=1&sn=52292535ca61af3fc66326f8aea772ce&chksm=ffdc0ae6d84a39877cf6ccc6357a241ab2bba9ccb47fda840fc078fdf1c84e2901fe23c51c07&scene=0&xtrack=1#rd

Video Credit: @Kling_ai on X

That’s all for today’s China AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.

The post China AI Native Industry Insights – 20261009 – Vidu | ima.copilot | Kling AI | more appeared first on AI Native Foundation.

1. SkyProduction launches Seedance 2.5 draft mode with 480P preview and 1080P final rendering to reduce AI video production costs

SkyProduction, powered by Kunlun Tech, launched the Seedance 2.5 draft mode on September 29, allowing AI video creators to generate 480P preview videos before committing to 1080P final renders. The feature addresses high inference costs in AI short drama production, where creators typically generate dozens of versions before selecting the final clip. According to testing from ByteDance’s Volcano Engine, the draft mode can reduce costs by approximately 57% in typical scenarios and up to 90% in high-iteration workflows, while maintaining visual consistency between draft and final output without relying on upscaling. The mode is now available on skyproduction.cn for users creating AI-generated video content.
Read more: https://mp.weixin.qq.com/s?__biz=MzI1MzE1NDc3Mg==&mid=2247510469&idx=1&sn=d9fe2fea61b8627bf8923d51d0d59933&chksm=e883c45632445d9212d3079b59b123de6d077faac06d24b893b833141199f0ba0e16604ce14e&scene=0&xtrack=1#rd

Video Credit: The original article

2. Tencent ima integrates multiple Tencent ecosystem Skills including Medical Knowledge, Stock Data, News, and WeChat Reading

Tencent announced that its AI assistant ima has integrated multiple Tencent ecosystem Skills, allowing users to access various services through voice commands without switching apps. The integrated Skills include Tencent Medical Knowledge Q&A, Tencent Stock Financial Data Query, Tencent News, WeChat Reading, Tencent Campus Recruitment, Tencent Fact Check, Tencent Meeting, Tencent E-Sign Contract Assistant, Tencent Weather, Tencent Channel, and PDF processing tools. Users can install these Skills through the Discover section in ima and invoke them directly through conversation, with outputs that can be saved to the knowledge base for future reference.
Read more: https://mp.weixin.qq.com/s?__biz=Mzk0MDg0MTkzOA==&mid=2247506155&idx=1&sn=45c22d07c0d94fede8cb10035c86bcb0&chksm=c3abee5afbd330c3f8c657471116c7a70f09ac825e871eab0f45c022161c9c91380d8ebfcdba&scene=0&xtrack=1#rd

Video Credit: NotebookLM

3. Qoder integrates TypeSafe AI’s Jev judgment model into Agent Harness for automated decision-making

Qoder announced its integration of Jev, a System One judgment model from TypeSafe AI, into its Agent Harness architecture to handle high-frequency semantic decisions in agent workflows. The integration separates responsibilities where generative models handle understanding and planning, Jev handles bounded semantic judgments, and Harness manages permissions, execution, and validation. Qoder demonstrated two use cases: Auto Mode, which uses Jev to determine whether agent operations can execute automatically with permission judgment latency reduced by approximately 83 percent from 3.13 to 0.53 seconds, and browser automation, where Jev selects next-step page operations within a controlled decision loop. The approach aims to establish a clearer division of labor between generation, judgment, deterministic rules, and user control in agent systems.
Read more: https://mp.weixin.qq.com/s?__biz=MzE5ODI2NzI5Nw==&mid=2247492846&idx=1&sn=171f168c0a3a970a06990ae6bfefa839&chksm=97bd8ffd0b4a074f9ea84bd84f2a40af8bbc4ed098f15f5fdeeee4e9381574323e96114b9358&scene=0&xtrack=1#rd

Video Credit: The original article

That’s all for today’s China AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.

The post China AI Native Industry Insights – 20260930 – Kunlun Tech | ima.copilot | Qoder | more appeared first on AI Native Foundation.

1. Alibaba Launches Qwen-Audio-3.1 With Five Upgraded Models and Major Price Cuts

Alibaba’s Qwen team released Qwen-Audio-3.1, upgrading its ASR, TTS, and Realtime audio models and adding two new models, TTS-Next and ASR-Next. The five-model lineup covers audio understanding, generation, interaction, and creation. ASR-Next adds multi-speaker recognition with speaker labels and timestamps plus sound and emotion understanding, while TTS-Next combines a language model with diffusion to generate voice, effects, and background audio in one pass. Alibaba also cut prices across the lineup, with TTS down about 70%, Realtime about 85%, and ASR up to 95% cheaper.
Read more: https://mp.weixin.qq.com/s/MW6wBdCJEdgGYX1DE9FzXQ

Video Credit: The original article

2. Kling announces Kling 4.0 video generation model with 4.0 Flash version available for early access

Kling announced that its upgraded Kling 4.0 video generation model will officially launch in October, with the Kling 4.0 Flash version opening for limited early access on September 28. The 4.0 version delivers major improvements in visual realism, creative control, and narrative capabilities, including enhanced camera movement handling, 4K 10-bit HDR output support, up to 30-second native video generation, multi-keyframe control with up to 10 frames, improved lip-sync accuracy, and support for up to 15 multimodal references including 10 images and 5 videos. The Flash version is designed for high-frequency creation scenarios with faster generation speed and cost efficiency, currently available to premium subscribers.
Read more: https://mp.weixin.qq.com/s?__biz=MzU5NTkwNDU2OA==&mid=2247496900&idx=1&sn=b0e629ecd3df2f3183f3585eef30a7f1&chksm=ffb9a9cfedc47b5cc650301acfd86535336dbfa10b49e35a848f25113284fbe66c9913db51b9&scene=0&xtrack=1#rd

Video Credit: The original article

3. SkyProduction launches AI MV creation, Mureka music models integration, and intelligent soundtrack features

On September 27, SkyProduction (Tiangong Workbench) released three new capabilities: AI MV creation tools, integration of Mureka V8, V9, and V9.5 music generation models into its infinite canvas, and intelligent soundtrack generation for videos and images. The AI MV feature analyzes song structure, rhythm, and lyrics to generate director plans and storyboards, then creates videos using the Seedance 2.5 model while maintaining character and scene consistency across shots. The music generation models allow users to create original songs from text prompts, while the intelligent soundtrack feature generates custom background music based on video or image content. All three features operate within a unified canvas workspace, enabling creators to generate music, create music videos, and add soundtracks without switching between different tools.
Read more: https://mp.weixin.qq.com/s?__biz=MzI1MzE1NDc3Mg==&mid=2247510445&idx=1&sn=986335f6da07fef6af3111ddf76db4bb&chksm=e871f887a0d891d5be2ac5c53b0e391289a46e0990e95944ded8f78c0a4b25588462eae54b8a&scene=0&xtrack=1#rd

Video Credit: The original article

That’s all for today’s China AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.

The post China AI Native Industry Insights – 20260929 – Alibaba | Kling AI | Kunlun Tech | more appeared first on AI Native Foundation.

1. PixVerse launches R2, a real-time world model for interactive story generation

PixVerse has introduced PixVerse R2, described as a new real-time world model. The system lets users explore generated worlds and control or edit them using text prompts. It supports shaping ongoing storylines and includes characters that can remember context and respond accordingly. The announcement was made via the company’s official account on September 22, 2026.
Read more: https://mp.weixin.qq.com/s/vjIPBLJXa_Plcbxidv_Qiw

Video Credit: @PixVerse_ on X

2. Alibaba’s Qwen Launches ‘Qwen Intelligence’ with Three SOTA Mobile Agents

Alibaba’s Qwen team introduced Qwen Intelligence, a personal AI system built around three specialized agents. The Mobile Planner Agent ranks first on MobilePA-Bench and its Business and Memory tracks for task planning and orchestration. The Mobile-Use Agent achieves 82.1 on MobileWorld, 92.2 on MobileWorld-Real, and 97.2 on AndroidDaily, with a 90% end-to-end task success rate, while the Mobile Creative Agent generates images in about 3 seconds, roughly twice as fast as leading competitors. Qwen also open-sourced its benchmark suite, including MobilePA-Bench, MobileWorld, MobileWorld-Real, and MobileWorld-Safety, covering planning, cross-app execution, real-device performance, and safety evaluation.
Read more: https://www.qwenintelligence.com/

Video Credit: The original article

3. Qoder Cloud Agents launches Batch feature for large-scale Agent task execution

Qoder Cloud Agents has released a Batch feature that enables large-scale, asynchronous execution of AI Agent tasks. The feature is designed for scenarios involving thousands of independent tasks that share similar processing logic, such as customer service quality checks, contract reviews, and Agent evaluation testing. Each batch supports up to 10,000 tasks submitted via JSONL files, with built-in state management, failure recovery, and result reconciliation. The platform offers off-peak scheduling between 22:00 and 08:00 Beijing time with promotional discounts on Qwen series models to reduce operational costs.
Read more: https://mp.weixin.qq.com/s?__biz=MzE5ODI2NzI5Nw==&mid=2247492649&idx=1&sn=535a3a994a1c8422804dfdf83f8eac8f&chksm=976b8e897a195e7ac5c1cf849ae3f549e2caa9144be2f4f023c12680a36c70861325ffa0c9d5&scene=0&xtrack=1#rd

Video Credit: NotebookLM

That’s all for today’s China AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.

The post China AI Native Industry Insights – 20260924 – PixVerse | Alibaba | Qoder | more appeared first on AI Native Foundation.

1. Alibaba Releases Qwen-Image-2.1 with Open Weights for Unified Image Generation and Editing

Alibaba’s Qwen team released Qwen-Image-2.1, a new open-weight model in the Qwen-Image series described as balanced and cost-effective. The model uses a lightweight 7B architecture that unifies image generation and editing, with natively supported RGBA transparency for layered compositing and text editing. It supports up to 10 reference images with high-fidelity local editing for portraits and products, and performs well on panoramas, infographics, and virtual try-ons. Model weights and code are available on Hugging Face, ModelScope, and GitHub.
Read more: https://www.modelscope.cn/models/Qwen/Qwen-Image-2.1

Video Credit: NotebookLM

2. Xiaomi Releases MiMo-V2.6 Pro and Flash Omnimodal AI Models

Xiaomi introduced MiMo-V2.6, a new family of omnimodal models comprising Pro and Flash versions, developed using scaled reinforcement learning. The company says the Pro model performs on par with Claude Opus 5 and GPT-5.6 Sol across most agent benchmarks and scores 46 on the Artificial Analysis Intelligence Index, the highest among open-source models. Xiaomi highlighted improvements in coding, computer use, 3D reasoning, and creative capabilities. The release includes open model weights, a technical report, RL environments, and training code.
Read more: https://mimo.xiaomi.com/mimo-v2-6

Video Credit: NotebookLM

3. Unitree Unveils Dex5-S Dexterous Robotic Hand with 22 Degrees of Freedom

Unitree Robotics introduced the Dex5-S, a precision biomimetic dexterous hand with 22 degrees of freedom and a size matching a real human hand. All 22 joints support smooth backdrivability and are equipped with limit impact torque protection. The hand is priced starting at $6,500, excluding tax and shipping. The announcement was made via Unitree’s official social media account.
Read more: https://mp.weixin.qq.com/s/RTl7NbTGo_IvvrqbaeI7nQ

Video Credit: @UnitreeRobotics on X

4. Tencent releases Hy Image 3.5 preview, a cost-effective image generation model for professional creative work

Tencent has officially launched Hy Image 3.5 preview, a professional-grade image generation model showing approximately 30% capability improvement over Hy Image 3.0 based on internal blind testing by hundreds of professional designers. The model supports text-to-image and image-to-image generation with up to 5 reference images, multiple aspect ratios, and up to 2K resolution output. It is integrated into multiple Tencent applications including Yuanbao AI assistant, WorkRally, OnSolo, Miora, WorkBuddy, and ima, and is available via Tencent Cloud API at a competitive price of 0.15 yuan per 2K image. The model demonstrates enhanced capabilities in text rendering and layout, realism, style expression, and editing consistency for both everyday creative tasks and professional production scenarios including commercial posters, film production, UI design, and e-commerce materials.
Read more: https://mp.weixin.qq.com/s?__biz=MzkwODU2OTQyNQ==&mid=2247498632&idx=1&sn=503321c4942df7d8983d81839f1292de&chksm=c1047ee7fcf9efd59761318565e538a470d85cb794b03cef1142ed8332bd951c46349b52c9bf&scene=0&xtrack=1#rd

Video Credit: NotebookLM

That’s all for today’s China AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.

The post China AI Native Industry Insights – 20260922 – Alibaba | xiaomi | Unitree | more appeared first on AI Native Foundation.