Where Did the BlueRun Ventures Family Take AI This Summer? | BlueRun Ventures Family Headlines

The AI frenzy is burning as hot as the summer itself. From the continuous evolution of AI agents to the release of various open-source models and the real-world deployment of embodied intelligence robots, we're witnessing AI's accelerating penetration across every domain — and seeing innovation move rapidly from the lab to the front lines of industry.

Today, we've mapped out the recent AI milestones from BlueRun Ventures portfolio companies to bring you a summer snapshot of the cutting edge.

Since August 20, Li Auto's VLA driver foundation model has begun customer delivery alongside the Li Auto i8, delivering the world's first VLA-based assisted driving system, with a full rollout starting September 10. Following the i8's launch, daily assisted driving usage tripled, cumulative mileage penetration increased 2.2x, and daily parking assistance usage rose 2.1x — evidence that the VLA driver model has become an assisted driving system users genuinely love and actually use.

On August 21, Li Auto's foundational model for its voice assistant, MindGPT 3.1, was officially released. As an intelligent agent language model with rapid inference capabilities, it features three key attributes: autonomous tool use, complex task completion, and personalized memory and understanding. In the agent era, the model is the agent. MindGPT will continue pushing technical boundaries for agent language models, autonomously planning decisions, calling tools, taking action, and solving problems to close the loop from cognition to execution.

On July 11, Moonshot AI officially released and open-sourced its Kimi K2 model. Kimi K2 is a Mixture-of-Experts architecture foundation model with stronger coding capabilities and greater proficiency at general agent tasks, featuring 1 trillion total parameters and 32 billion active parameters.

On benchmark tests including SWE-Bench Verified, Tau2, and AceBench, Kimi K2 achieved SOTA results among open-source models, demonstrating leading capabilities in coding, agent tasks, and mathematical reasoning. Kimi K2 uses the MuonClip optimizer to robustly support trillion-parameter model training, significantly improving token efficiency. Combined with large-scale agentic data synthesis and general reinforcement learning, the model continues to advance in general intelligence capabilities.

The Kimi K2 release also drew significant global attention. On July 16, the website of Nature, one of the world's top scientific journals, published an article praising the model release as "another DeepSeek moment."

Try it out 👇

$2

Model and FP8 weights download links 👇

$2

In April this year, Genspark Super Agent officially launched. The agent employs a Mixture-of-Agents (MoA) system integrating 9 mainstream models and 80 tools, capable of understanding user intent, autonomously planning and executing tasks, and supporting multimodal processing including travel planning, video creation, and academic research. On the authoritative GAIA benchmark, Genspark Super Agent topped the leaderboard with a 98.5% task completion rate, becoming the world's first full-featured AI Agent available without an invite code.

Since launch, Genspark Super Agent has successively introduced multiple sub-agents to enrich its built-in functionality:

• Genspark AI Call: Supports a complete closed loop from intent understanding to phone execution. Users simply input natural language instructions to deploy a voice interaction sub-agent that handles phone bookings, business communications, and other tasks, calling text-to-speech (TTS), automatic speech recognition (ASR) models, and telecom APIs to complete the loop;
• Genspark AI Docs: Simply tell Genspark your needs and watch a complete, professionally designed document materialize within minutes. This is also the world's first and only native AI document tool supporting rich text and Markdown formats, with instant conversion between formats;
• Genspark AI Pods: With simple user instructions, the AI system immediately processes the request through content analysis, research processing, audio production, and host generation, delivering a complete, professional-grade podcast in minutes;
• Genspark AI Meeting Notes: The world's first AI meeting notes tool for Apple Watch. Through tight integration with Genspark Super Agent, users can convert any meeting into comprehensive, professional notes with a double-tap on their Apple Watch or a single tap on their phone;
• Genspark AI Drive: The world's first complete agent download agent seamlessly integrated with the all-new Genspark AI Drive. With a single command, users can download, organize, and interact with any file, collecting, organizing, and managing web content with remarkable speed and accuracy;
• Genspark AI Developer: A fully automated programming agent supporting the complete pipeline from requirements analysis to code generation and deployment. Users can describe needs in natural language and generate complete projects with frontend interfaces, backend management systems, and payment gateways in 10 minutes, with GitHub version control integration;
• Genspark AI Designer: By integrating text semantic encoders with visual feature extractors, completes full brand design suites, supporting one-stop generation from logos and print materials to product packaging, websites, apps, and social media advertising — removing design barriers so anyone can build their vision without technical constraints;
• Genspark Clip Genius: From podcast summaries and sports clips to gaming montages and personal vlogs, Genspark Clip Genius handles intelligent content analysis, clever story design, precise editing and optimization, and professional post-production integration — editing any video from a single prompt and automatically completing the process from analysis to final render in minutes.

Try Genspark Super Agent now 👇

$2

On June 10, Genspark launched another disruptive product: a fully agentic AI browser. Genspark browser embeds AI directly into every webpage, giving the browser itself the ability to think and act, connecting various commonly used tools and applications to form automated workflows — providing shopping price comparisons, video summaries, Autopilot Mode, and even letting the browser itself browse Weibo, organize information, and create personal podcasts.

What should a browser look like in the AI era? Genspark's attempt offers an intriguing answer: a browser shouldn't just be a "window," but a true intelligent assistant.

In August, AgiBot unveiled Genie Envisioner, a unified world model platform for real-world robot manipulation. Unlike traditional fragmented "data-training-evaluation" pipelines, GE integrates future frame prediction, policy learning, and simulation evaluation into a video-generation-centric closed-loop architecture for the first time, enabling robots to complete end-to-end perception, reasoning, and execution within a single world model — from "seeing" to "thinking" to "acting."

On July 7, AgiBot released the Lingxi X2-N robot. One of its standout features is the ability to freely switch between bipedal and wheeled modes. In bipedal mode, the X2-N can navigate stairs blind while carrying 6 kg loads; in wheeled mode, it handles complex terrain including single-plank bridges and slopes.

During the 2025 WRC, the world's first city-level, permanently operating humanoid robot demonstration zone officially launched in Haidian, with Galaxy Universal's humanoid robot smart retail solution tailored for urban life — the Galaxy Space Pod — making its debut.

Operated by Galaxy Universal's self-developed Galbot robot, the Galaxy Space Pod completes voice reception, order and payment processing, precise grasping, and product delivery, achieving "zero teleoperation, fully autonomous" operation. Each pod requires as little as 9 square meters of space yet operates autonomously 24/7, selling high-frequency consumer goods including beverages, snacks, cultural creative products, and pharmaceuticals.

In June, Galaxy Universal released GroceryVLA, the world's first end-to-end embodied VLA foundation model for the retail industry, marking a major technical breakthrough in embodied intelligence. GroceryVLA's five core capabilities enable stable adaptation to complex supermarket environments and highly general autonomous operation: strong applicability, high generality, cross-scene generalization, autonomous decision-making, and strong anti-interference.

On August 5, Hefei Youibot Intelligent Robot Co., Ltd. and the Youibot–Xi'an Jiaotong University Embodied Intelligent Robot Research Institute jointly released Lingxu, the world's first "one brain, multiple forms" cross-scene cluster-operation humanoid robot.

As the bipedal humanoid robot in the research institute's "Tianyan" embodied intelligent robot matrix, Lingxu is built on a "one brain, multiple forms" embodied intelligence foundation model architecture. Designed for the high-precision, long-process, multi-task compound generalization demands of semiconductor manufacturing and energy industries, it uses the robot intelligent brain MAIC (Mobile AI Comprehension) as its core to achieve efficient cluster coordination with robots of multiple forms, enabling multimodal fused perception, adaptive multi-arm collaborative operation, multi-form mobile control, and global logistics scheduling — leading industrial embodied intelligence into the era of cross-scene, highly generalized cluster operations.

On August 27, Haivivi launched the CocoMate series, the world's first AI toy powered by an end-to-end voice model, marking a new stage for the industry's technology. Compared to BubblePal, CocoMate achieves dual breakthroughs in hardware and interaction:

• Technical innovation: CocoMate leverages a new end-to-end voice model to achieve direct voice-to-voice output, maximally preserving user emotion and context while possessing anthropomorphic emotional capabilities for warm, genuine companionship;
• Enhanced interaction: CocoMate supports natural voice dialogue and multiple wake-up methods including shake and voice, with customizable wake words;
• Design and form factor: CocoMate features a detachable module that embeds into plush toy bodies, facilitating washing and skin swapping. The launch lineup includes Ultraman IP limited editions of Tiga and Zero, plus Haivivi's original IP "Bubble," with more characters to follow.

Recently, TingLixiong AI foundation model scientist Chi Zhang's team published research on a video production tool — FlowDirector: a completely new training-free video editing framework.

Two key technical breakthroughs:

1. Localized editing (SAFC): Traditional AI editing methods are like covering your entire body in paint — one slip and you're disfigured. The FlowDirector developed by Chi Zhang's team comes with an intelligent beauty mask that makes only your desired changes visible while automatically locking the rest of the background.

2. Smart eraser (DAG): Like sketching with pencil first, then erasing excess lines. The system first generates a "modification draft," then automatically erases parts overlapping with the original video, leaving only the perfect final edit.

In August, TingLixiong also launched the AI Companion Phone X3, China's first AI phone deeply customized for teenagers. Beyond basic functions like communication, camera, payment, and social features, it helps parents remotely monitor and ensure children's safety at all times. The device also includes multiple AI tools to accompany children's growth from all angles, around the clock.

Luchen Tech launched Video Ocean Agent, the world's first video agent integrated with GPT-5. Video Ocean Agent doesn't merely accelerate a single point in the process — it seamlessly connects the entire AI production pipeline into a "cloud-based creative team."

• Script planning module — understands a single creative instruction and automatically breaks down shots and pacing;
• Visual synthesis module — retrieves or generates dynamic and static assets based on the script, matching color schemes and subtitle styles;
• Voiceover and subtitle module — synthesizes multilingual background music and sound effects according to character and context, with real-time subtitle embedding;

Click to experience 👇

$2

On June 12, LLVision officially launched its consumer-grade AR glasses Leion Hey2 in Seoul, South Korea. The product achieves a "perception-free" real-time translation experience at interpreter level, breaking through the industry "impossible triangle" of lightweight design, performance, and battery life. The complete device weighs only 49 grams, supports real-time translation for 100+ languages with under 500ms latency, delivers 8 hours of standalone operation, and achieves 96 hours of battery life with the portable glasses case — suitable for cross-border meetings, overseas travel, study abroad classrooms, and diverse other scenarios.

Since beginning testing on July 21, Fabarta's personal dedicated intelligent agent has been applied to vertical domain office writing scenarios including law, education, and government affairs. On August 4, beyond optimizing and strengthening the existing "Professional Edition," Fabarta launched "Community Edition" and "Enterprise Edition" versions, precisely adapting to different user groups and enterprise-level application environments. All three versions include core functions like intelligent writing and knowledge Q&A.

For professional individual users, Fabarta's personal dedicated intelligent agent enables a complete writing logic chain of "local knowledge application — intelligent framework generation — professional content verification." Meanwhile, every generated draft is沉淀为 local knowledge, reducing the cost of subsequent writing.

At the organizational level, Fabarta's personal dedicated intelligent agent fully connects cloud, edge, and local environments, achieving integration of enterprise knowledge bases, data services, and local knowledge bases. It串联集成 employees' personal intelligent agents with enterprise knowledge centers, enterprise data, and business systems, embedding AI agent capabilities into enterprise workflows to improve organizational efficiency.

On June 24, Jobright officially launched Jobright Agent, becoming the first AI career agent that finds, customizes, and submits applications on behalf of job seekers.

Jobright Agent reshapes every aspect of the job search experience, transforming the model from manual passive processing to intelligent proactive career matching:

• From search to matching: Traditional platforms require users to manually sift through massive job listings. Jobright's intelligent agent matches opportunities based on skills, experience, and career goals;
• From manual to automatic: Traditional job sites force users to manually filter positions and repeatedly submit applications, while Jobright Agent automates these tedious processes so job seekers can focus on truly valuable activities like interview preparation;
• From one-way matching to mutual fit: Going beyond keyword matching, Jobright Agent evaluates mutual fit, significantly improving hiring success rates for both candidates and recruiting teams;
• From passive waiting to proactive outreach: Rather than passively waiting for users to discover opportunities, Jobright Agent actively searches for relevant positions and initiates applications on the job seeker's behalf.

Jobright is trusted by over 520,000 professionals and growing 30% year-over-year. Jobright Agent delivers 2x more interviews for job seekers and reduces job search time by 80%.

Intelligent Surge: BlueRun's AI Portfolio Drops Even More New Products! | BlueRun Ventures Family Headlines

Stop Asking Me What an AI Agent Is | BlueRun x Volcano Engine Event Recap BlueRun Ventures Portfolio Companies Give WAIC a Taste of AI New Species

**