1. Xiaomi releases and open-sources MiMo Code V0.1.0, an AI programming assistant with persistent memory system

Xiaomi has released and open-sourced MiMo Code V0.1.0, a terminal-based AI programming assistant built on the OpenCode project under MIT license. The tool features a persistent memory system that addresses long conversation context loss through project memory, session checkpoints, and task progress tracking mechanisms. It includes the MiMo-V2.5 multimodal model with performance comparable to Claude Sonnet 4.6, supports integration with mainstream models like DeepSeek, Kimi, and GLM, and introduces a Compose mode for end-to-end software development workflows. The system achieved 62% on SWE-Bench Pro and 73% on Terminal Bench 2 benchmarks, outperforming Claude Code using the same underlying model.

Read more: https://mp.weixin.qq.com/s/WkMPz-eBK2Hz0ZTQEwtx6w

Video Credit: The original article

2. MNN inference engine adds Arm SME2 support achieving 80% speedup for Qwen3-VL on-device deployment

Alibaba’s MNN inference engine has integrated support for Arm’s second-generation Scalable Matrix Extension (SME2) instruction set, enabling significant performance improvements for on-device AI model deployment. Testing with the Qwen3-VL-4B-Instruct multimodal model on SME2-enabled flagship devices like vivo X300 showed an 81% speedup in the prefill stage and 13% improvement in the decode stage compared to non-SME2 implementations. The integration uses a compile-time switch with runtime hardware detection, automatically selecting the optimal acceleration path without requiring manual configuration. MNN has released complete deployment tools including model conversion, quantization, and Android app integration capabilities for developers.

Read more: https://mp.weixin.qq.com/s/QNl4pn5JzxpEeFAEvq88ZA

Video Credit: The original article

3. Tencent Hunyuan open-sources HPC-Ops inference operator library with major system-level upgrade

Tencent Hunyuan AI Infra team has released a major upgrade to its open-source HPC-Ops inference operator library, expanding from individual operators to a comprehensive optimization suite covering the entire inference pipeline. The update includes five key operators: dynamic load-balanced Attention that achieves up to 2.95x speedup on long-context workloads and 17% QPM improvement, Router GEMM using dual BF16 for FP32-level precision with up to 3.22x speedup over cuBLAS FP32, FusedMoE with 1.2x to 1.6x performance gains over vLLM and SGLang, Fused AllReduce+Norm delivering 1.04x to 1.68x speedup, and Sampler achieving 4.0x to 7.5x speedup over vLLM. The library is production-ready, fully open-source, and available immediately on GitHub for developers building high-throughput, low-latency large model inference services.

Read more: https://mp.weixin.qq.com/s/y0vd1cpvWXLSssL1kXxguw

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.

1. Anthropic releases Claude Fable 5 and Claude Mythos 5 with enhanced capabilities and safeguards

Anthropic announced Claude Fable 5, the first publicly available Mythos-class model featuring state-of-the-art performance in software engineering, knowledge work, and vision tasks. The model includes built-in safeguards that block responses in high-risk areas like cybersecurity and biology, instead routing to Claude Opus 4.8 for safety. Concurrently, Anthropic launched Claude Mythos 5 for trusted partners in the Glasswing program, providing unrestricted access to advanced capabilities for cybersecurity professionals. Both models represent a significant jump in capability over previous versions and are priced at $10 per million input tokens and $50 per million output tokens.

Read more: https://www.anthropic.com/news/claude-fable-5-mythos-5

Video Credit: Claude Youtube Channel

2. Google releases DiffusionGemma experimental open model with 4x faster text generation using diffusion-based parallel processing

Google released DiffusionGemma, an experimental open model under Apache 2.0 license that uses text diffusion to generate entire blocks of text simultaneously rather than token-by-token. The 26B Mixture of Experts model activates only 3.8B parameters during inference and delivers up to 4x faster text generation on dedicated GPUs, achieving speeds of 1000+ tokens per second on NVIDIA H100 and 700+ tokens per second on RTX 5090. The model is designed for researchers and developers working on latency-sensitive workflows including code infilling, inline editing, and non-linear text generation. Google notes that DiffusionGemma prioritizes speed over quality and performs below standard Gemma 4 on benchmarks.

Read more: https://blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/

Video Credit: @googlegemma on X

3. Google releases Gemini 3.5 Live Translate for real-time speech translation in 70+ languages

Google announced the release of Gemini 3.5 Live Translate, a new audio model that provides live speech-to-speech translation in over 70 languages. The model generates continuous translated speech while preserving speaker intonation, pacing, and pitch, staying just seconds behind the speaker throughout sessions. The technology is rolling out to Google Translate on Android and iOS, Google Meet with expanded language support, and is available to developers through the Gemini Live API and Google AI Studio. All generated audio includes SynthID watermarks for AI content detection.

Read more: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-live-3-5-translate/

Video Credit: @GoogleAI on X

4. Cursor announces Bugbot code review agent is 3x faster, 22% cheaper, and finds 10% more bugs

Cursor released major improvements to its Bugbot code review agent, making it over 3x faster to run, 22% cheaper, and capable of finding 10% more bugs per review. The company reports that 90% of Bugbot runs now finish in under three minutes. Users can also access a new /review command to run Bugbot locally to catch and fix issues before pushing code to repositories.

Read more: https://cursor.com/blog/bugbot-updates-june-2026

Video Credit: @cursor_ai on X

That’s all for today’s Global 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.

1. Huawei Cloud launches Agentic Infra paradigm and AI products at INSPIRE 2026 conference

Huawei Cloud unveiled Agentic Infra, a new infrastructure paradigm for general and AI workloads at its INSPIRE 2026 conference in Shanghai. The company released four core products including AI Cluster Service supporting over 100,000 cards with 200 EFLOPS computing power, Agentic Memory Storage with petabyte-scale memory space, ModelArtsNext training platform, and AgentArts enterprise agent platform. The new infrastructure features efficient token generation, continuous learning, unified compute scheduling, and secure autonomy for enterprise AI deployment.

Read more: https://www.huawei.com/en/news/2026/6/inspire-agenticera-agenticinfra

Video Credit: NotebookLM

2. Xiaomi MiMo-V2.5-Pro-UltraSpeed achieves breakthrough of 1000 tokens per second inference speed for 1T parameter model

Xiaomi and TileRT jointly announced the MiMo-V2.5-Pro-UltraSpeed mode, marking the first time a trillion-parameter model has exceeded 1000 tokens per second in output speed. The breakthrough was achieved through co-design optimization including FP4 quantization for MoE experts and DFlash speculative decoding with block-level masked parallel prediction, running on standard 8-GPU nodes. The UltraSpeed API is now available through an application-based limited release from June 9 to June 23, 2026, priced at 3x the standard MiMo-V2.5-Pro rate while delivering approximately 10x faster output. Over 3000 enterprises and developers from various industries including law, finance, logistics, and automotive manufacturing applied for trial access within 12 hours of the announcement.

Read more: https://mp.weixin.qq.com/s/EZvmrx8xfM9MZNCMDwImFQ

Video Credit: The original article

3. Kimi Code releases major upgrade with one-command installation, video understanding, and ACP protocol support

Moonshot AI announced a major version upgrade for Kimi Code, its open-source Coding Agent product. The update introduces one-command installation with millisecond-level startup, video understanding capabilities including color grading extraction and video editing, and integration with authoritative data sources such as TongHuaShun and Tianyancha for real-time stock prices and financial reports. The upgrade also adds support for the ACP protocol, enabling use within JetBrains and Zed editors, and includes a comprehensive hook ecosystem for integration with other tools. Users can access the updated version at kimi.com/code.

Read more: https://mp.weixin.qq.com/s/sqGuRyeU9AvZZ3IXj8ddJg

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.

1. Cognition launches FrontierCode benchmark to evaluate AI code mergeability beyond correctness

Cognition introduces FrontierCode, a new AI coding benchmark developed with 36 open-source maintainers who each invested over 40 hours per task. Unlike traditional benchmarks that focus on functional correctness, FrontierCode measures whether AI-generated code would actually be accepted by human maintainers for production codebases. The benchmark evaluates code along six axes including behavioral correctness, regression safety, mechanical cleanliness, test correctness, scope discipline, and code quality. Initial results show that over half of outputs passing earlier SWE-Bench tests fall short on mergeability standards, with even advanced models like Anthropic’s Claude Opus 4.8 scoring only 13.4% on the most difficult Diamond tasks.

Read more: https://cognition.ai/blog/frontier-code

Video Credit: NotebookLM

2. Google upgrades NotebookLM with agentic capabilities, Gemini 3.5, and new output formats

Google announced major upgrades to NotebookLM that add agentic AI capabilities, advanced reasoning with Gemini 3.5, and ability to generate multiple file formats. The tool can now autonomously search for sources, write and run code through a secure cloud computer, and create downloadable outputs like PDFs, spreadsheets and presentations. The upgrade transforms NotebookLM from a document summarizer into an autonomous research assistant that can handle complex multi-step research projects. The new capabilities are rolling out to Google AI Ultra subscribers and Workspace customers with expanded access.

Read more: https://blog.google/innovation-and-ai/products/notebooklm/better-research-notebooklm/

Video Credit: @NotebookLM on X

3. Apple expands Private Cloud Compute to Google Cloud using NVIDIA GPUs for Apple Intelligence workloads

Apple announced it is expanding Private Cloud Compute beyond its own data centers to Google Cloud, using NVIDIA GPUs to run new Apple Intelligence workloads. The collaboration extends Apple’s privacy protections to third-party data centers for the first time, utilizing NVIDIA Confidential Computing with NVIDIA GPUs, Intel CPUs with TDX, and Google’s Titan chip. Apple worked with Google to leverage Gemini model technologies for building next-generation Apple Foundation Models, with the most demanding tasks including agentic tool-use and complex reasoning running on NVIDIA GPUs in Google Cloud while maintaining Apple’s security and privacy protections.

Read more: https://security.apple.com/blog/expanding-pcc/

Video Credit: NotebookLM

That’s all for today’s Global 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.

1. Alibaba TongyiLab releases AgentScope 2.0 multi-agent framework with system-level transparency features

Alibaba’s TongyiLab announced AgentScope 2.0, a production-ready agent framework focused on system-level transparency for multi-agent applications. The new version includes built-in retry and fallback mechanisms, an observable event system, a smarter permission system, and decoupled workspace with unified agent service. The framework is designed to work with increasingly capable LLMs while maintaining complete visibility into agent operations and decision-making processes.

Read more: https://github.com/agentscope-ai/agentscope

Video Credit: @Ali_TongyiLab on X

2. ResNet paper co-authored by StepFun Chief Scientist Zhang Xiangyu receives CVPR 2026 Longuet-Higgins Prize

CVPR 2026 announced that the 2015 paper Deep Residual Learning for Image Recognition (ResNet), co-authored by Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, has received the Longuet-Higgins Prize. The prize recognizes research that has demonstrated long-term impact on both academic research and industrial development in computer vision. ResNet introduced residual learning to solve deep neural network training challenges and has become a foundational architecture in modern deep learning, with over 320,000 citations making it the most cited paper of the 21st century. The residual connection concept has expanded beyond computer vision to natural language processing, speech, multimodal systems, and other AI domains.

Read more: https://mp.weixin.qq.com/s/ZVgqdH_fE42jO4kcI-lF3g

Video Credit: Hyperframes

3. Tencent Hunyuan proposes Stem sparse attention algorithm reducing time-to-first-token by 3.6x, paper accepted to ICML 2026

Tencent Hunyuan announced the Stem sparse attention algorithm, which has been accepted to the machine learning conference ICML 2026. The algorithm introduces Token Position Decay (TPD) and Output-Aware Metric (OAM) innovations to achieve near-dense attention accuracy using only 25% of computational resources. When integrated into the Hunyuan Hy3 preview model with optimized HPC operators, Stem reduces time-to-first-token by 3.6x for 128K context lengths. The algorithm addresses bottlenecks in long-context inference by reallocating computational budget based on causal information flow and evaluating token importance through both attention scores and value vector magnitudes. Both the Stem algorithm and HPC operator implementations have been open-sourced on GitHub.

Read more: https://mp.weixin.qq.com/s/XneOSvjt-7A-DU546cGoZA

Video Credit: Hyperframes

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.

1. Astra Autonomous Pentest

Product Hunt Data
Ranking: 11
Upvote: 416

Product Overview
Astra Autonomous Pentest is an agentic security system that continuously runs offensive testing to discover chained vulnerabilities, validates findings to minimize false positives, and then generates code-ready remediation guidance as native prompts for developer tools like Cursor, Copilot, and Claude Code, turning pentesting into an always-on loop rather than a point-in-time report.

Evaluation
AI Native Application Modernization: 89/100
The product is strongly AI-native because multi-agent discovery, independent validation, and fix generation are the core workflow, not add-ons; it modernizes security operations by connecting detection to developer execution via IDE-native prompts, though real-world impact will depend on how well it fits into diverse SDLCs and governance requirements.

Website
https://www.getastra.com

2. Ideogram 4.0

Product Hunt Data
Ranking: 29
Upvote: 252

Product Overview
Ideogram 4.0 is an open-weight text-to-image model built for production-grade visual generation, combining prompt-to-image creation with bounding-box layout control, reliable multilingual text rendering, and native 2K outputs for design workflows and developer integrations.

Evaluation
AI Native Application Modernization: 86/100
The core system is an AI model trained from scratch and exposed as a controllable generation engine, enabling modern app patterns like programmatic layout constraints, consistent typography in images, and high-resolution output; modernization is strong, with deployment and governance effort still required for enterprise adoption.

Website
https://ideogram.ai/models/4.0

3. Spectron

Product Hunt Data
Ranking: 45
Upvote: 180

Product Overview
Spectron is an AI-agent memory substrate that unifies vectors, graph, documents, and relational-style rows under a single ACID transaction model, so agent knowledge can be written, corrected, and retrieved without cross-store sync. It preserves provenance per fact, supports corrections that supersede rather than overwrite, and uses hybrid retrieval with trace-informed ranking to improve agent recall and grounding.

Evaluation
AI Native Application Modernization: 86/100
Spectron is strongly AI-native because memory, provenance, correction semantics, and hybrid retrieval are core primitives designed around agent workflows rather than a conventional database with add-on embeddings. The single transactional substrate reduces operational complexity and inconsistency risk, while tri-temporal facts and tenant scoping align with production agent requirements; the main adoption work is mapping existing data models and retrieval stacks onto its unified approach.

Website
https://surrealdb.com

4. Nemotron 3 Ultra by NVIDIA

Product Hunt Data
Ranking: 49
Upvote: 6

Product Overview
Nemotron 3 Ultra is an open-weights, large-scale Mixture-of-Experts reasoning model designed to run long-horizon agent loops with high throughput and an unusually large context window, and it can be deployed via common model hubs and NVIDIA NIM as an inference microservice for production agent systems.

Evaluation
AI Native Application Modernization: 89/100
This is strongly AI-native because the model is the core execution layer for agent reasoning, with architecture and deployment packaging optimized for multi-step workflows, long context, and scalable serving; the main modernization gap is that teams still need surrounding orchestration, tool execution, and governance to turn raw model capability into end-to-end applications.

Website
http://www.nvidia.com

Statement: Evaluation results are generated by AI, lack of data support, reference learning only.

1. Anthropic reports Claude now authors 80% of company’s production code as AI development accelerates toward potential recursive self-improvement

Anthropic announced that its AI model Claude now authors more than 80% of the code merged into the company’s production systems as of May 2026, up from single digits before February 2025. The company’s internal data shows this represents a possible path toward recursive self-improvement, where AI systems could autonomously build more capable successors. Anthropic engineers now ship 8 times as much code per quarter compared to 2021-2025, with Claude handling coding tasks that previously took humans hours to complete. The company warns this trend toward AI systems accelerating their own development is happening faster than expected and deserves greater attention from policymakers.

Read more: https://www.anthropic.com/institute/recursive-self-improvement

Video Credit: NotebookLM

2. OpenAI launches upgraded ChatGPT memory system with dreaming architecture for enhanced context retention

OpenAI announced a significant upgrade to ChatGPT’s memory capabilities through a new dreaming architecture that synthesizes information from past conversations automatically in the background. The system addresses previous issues with memory staleness and accuracy by creating a more capable and compute-efficient memory infrastructure. The upgrade includes a memory summary page where users can review what ChatGPT knows about them and manage personalization settings. Plus and Pro users in the US receive doubled memory capacity starting today, while free users will gain access to dreaming-based memory features in coming weeks due to 5x compute efficiency improvements.

Read more: https://openai.com/index/chatgpt-memory-dreaming/

Video Credit: @OpenAI on X

3. OpenAI launches Build iOS Apps plugin for Codex with SwiftUI preview and hot reload features

OpenAI released the Build iOS Apps plugin for Codex, allowing developers to view and test iOS apps in the in-app browser, open SwiftUI previews, and hot reload edits without leaving Codex. The plugin provides skills for iOS debugging, performance analysis, and SwiftUI development workflows. It integrates with XcodeBuildMCP to enable build, run, and debug operations directly through Codex’s interface, eliminating the need to switch to Xcode’s GUI during development.

Read more: https://x.com/i/web/status/2062599291479478275

Video Credit: @OpenAIDevs on X

That’s all for today’s Global 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.

1. OpenAI launches Sites feature for Codex to create interactive hosted websites and apps

OpenAI announced Sites, a new Codex feature that allows users to create, share, and host interactive websites and apps through natural language prompts. Sites is rolling out in preview to Business and Enterprise customers first, before expanding more broadly. Users can transform ideas, analysis, and plans into dashboards, planners, review workspaces, and other interactive tools that can be shared with workspace members via URL.

Read more: https://openai.com/index/codex-for-every-role-tool-workflow/

Video Credit: @OpenAI on X

2. Google releases Gemma 4 12B, a unified encoder-free multimodal model for laptops

Google DeepMind released Gemma 4 12B, a 12-billion parameter multimodal model that processes text, images, and audio without separate encoders. The model runs locally on laptops with just 16GB of VRAM or unified memory and is released under an Apache 2.0 license. It features native audio processing capabilities as the first mid-sized Gemma model to support audio input, and nearly matches the performance of the larger 26B model across benchmarks while using less memory.

Read more: https://blog.google/innovation-and-ai/technology/developers-tools/introducing-gemma-4-12b/

Video Credit: Google Youtube Channel

3. Microsoft announces OpenClaw support for Windows with security containers and enterprise controls

Microsoft announced a partnership with OpenClaw to bring the open-source AI agent framework to Windows and enterprise environments. OpenClaw now runs securely on Windows using Microsoft Execution Containers (MXC) technology, providing policy-driven containment for enterprise deployments. The integration includes a Windows companion app for easy setup and connection to existing OpenClaw agents, along with Agent 365 integration for enterprise governance and security controls.

Read more: https://blogs.windows.com/windowsdeveloper/2026/06/02/windows-platform-security-for-ai-agents/

Video Credit: Microsoft Developer Youtube Channel

That’s all for today’s Global 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.