Vol.40 · July 20, 2026
dera news AI Weekly Vol.40 | 2026-07-20 - This Week's AI News
🤖 dera news AI Weekly Vol.40
2026-07-20
This week's AI world in one sentence? China's Moonshot AI releases Kimi K3, the first open 3T-class MoE model, accelerating the era of self-hosted AI. This 2.8 trillion parameter model, with MXFP4 quantization, drastically reduces storage, bringing cutting-edge performance to developers and offering a powerful alternative to existing closed models.
📊 What You Need to Know This Week
This week, the AI industry buzzed with the evolution of open-source models and the new possibilities they bring. Of particular note is Moonshot AI's announcement of Kimi K3, the first open 3T-class MoE model. Despite its massive 2.8 trillion parameters, MXFP4 quantization technology significantly reduces its storage requirements, offering developers a realistic option for operating state-of-the-art AI models in-house. This development could enable businesses previously reliant on closed-source models to adopt more flexible and cost-efficient AI solutions.
Concurrently, Thinking Machines' release of Inkling, an open-weight MoE model designed for customization, is good news for businesses looking to build AI tailored to specific tasks. Furthermore, PrismML's Bonsai 27B showcased a groundbreaking advancement by enabling a 27B-class multimodal AI to run on an iPhone through 1-bit quantization. This greatly expands the potential for cloud-independent, on-device AI agents, holding significant implications for SMBs prioritizing data security and real-time processing.
Meta's commercial release of DINOv3, an image recognition model leveraging self-supervised learning (SSL), also marks a move to lower AI adoption barriers. By enabling high-precision image recognition without the need for labeling, it could accelerate AI integration in diverse visual-information-critical tasks like product inspection and environmental monitoring. Within the Asian AI ecosystem, NVIDIA Nemotron is gaining traction, with Sakana AI, SoftBank, NTT, and NEC adopting it to accelerate the development of AI tailored to Japan's unique language and industrial needs.
On the other hand, the introduction of the "Long-Horizon-Terminal-Bench" AI agent evaluation standard highlights the challenges in measuring true AI capability and the room for improvement in complex, long-duration tasks. In terms of infrastructure, the "MCP" for connecting AI models with external tools will become stateless in its 2026 revision, improving scalability and making large-scale AI system deployment easier. However, industry friction is also evident; Apple's lawsuit against OpenAI for trade secret theft underscores the deteriorating relationship between the former collaborators and the growing importance of intellectual property protection in AI. Additionally, Google's Gemini 3.5 Pro release delay, particularly due to concerns about its code generation capabilities compared to competitors, signals the fierce nature of AI development competition.
What we're watching closely is how open-source AI is closing the performance gap with frontier models while enhancing adaptability for diverse devices and applications. This trend is accelerating the democratization of AI, creating opportunities for more businesses to integrate AI into their problem-solving strategies.
💡 This Week's Actions
1. Explore the potential of frontier open-source AI with Kimi K3 (30 minutes) Moonshot AI's Kimi K3 offers cutting-edge performance as an open-weight model. As a first step toward considering self-hosted AI as an alternative to expensive closed-source models in your AI strategy, thoroughly review its technical specifications and benchmark results. → Kimi K3, First Open 3T-Class MoE Model, Arrives
2. Experiment with new on-device AI experiences using Bonsai 27B (1 hour) Bonsai 27B, running on an iPhone, enables secure AI utilization without the cloud. Download the weights available on Hugging Face and try a demo of an AI agent on your iPhone to experience firsthand the potential of on-device AI for business automation and customer support. → Bonsai 27B: 27B Multimodal AI Runs on iPhone
3. Consider leveraging customizable open models for specialized AI (1.5 hours) Open-weight models like Thinking Machines' Inkling and NVIDIA Nemotron can be fine-tuned with your own data to build AI optimized for specific business tasks. Identify your business challenges, explore how these models can contribute to solutions, and plan a Proof of Concept (PoC). → Thinking Machines Ships Inkling, First Open-Weight MoE → NVIDIA Nemotron Gains Traction in Japanese AI Ecosystem
📰 This Week's AI Articles (10 stories)
1️⃣ Kimi K3, First Open 3T-Class MoE Model, Arrives
🏷️ AI Models, Open Source, Large Language Models What happened? China's Moonshot AI announced the release of Kimi K3, its first open 3T-class Mixture-of-Experts (MoE) model. This 2.8 trillion parameter model uses MXFP4 quantization to reduce storage to approximately 1.4TB, making self-hosted deployment a realistic option. It has shown performance surpassing Claude Fable 5 in frontend code arenas, pushing the frontier of open-weight AI. Our take The arrival of Kimi K3 suggests that open-source AI is now closing in on the performance of state-of-the-art closed-source models. The significant size reduction via MXFP4 quantization, in particular, opens the door for SMBs and developers to operate high-performance AI more affordably and securely within their own environments. This is a crucial step in democratizing AI and has the potential to reshape the industry's competitive landscape. 📎 Read more
2️⃣ Thinking Machines Ships Inkling, First Open-Weight MoE
🏷️ AI Models, Open Source What happened? Mira Murati's lab, Thinking Machines, released Inkling, its first open-weight foundation model, under an Apache 2.0 license. This Mixture-of-Experts (MoE) model boasts 975 billion total parameters, a 1-million-token context window, and multimodal inference capabilities. It is explicitly designed as a fine-tuning base for their Tinker customization platform, positioning it as a foundation for enterprises to build specialized AI. Our take Inkling is groundbreaking not just for its high performance, but for its design philosophy centered on "being customized." This makes it an incredibly powerful tool for SMBs looking to efficiently develop AI tailored to specific tasks or industries. The integration with the Tinker platform offers developers a new avenue to optimize models with their own data and establish a competitive edge. 📎 Read more
3️⃣ Bonsai 27B: 27B Multimodal AI Runs on iPhone
🏷️ AI Models, Multimodal AI, Open Source What happened? PrismML successfully compressed its 27B-class multimodal AI model, Bonsai 27B, to just 3.9GB using 1-bit quantization, enabling it to run on an iPhone 17 Pro. This allows advanced reasoning and visual tasks to be performed on-device without cloud connectivity. The model retains approximately 90% of the full-precision model's performance across 15 benchmarks, marking a significant step towards practical on-device agents. Our take Bonsai 27B represents a revolutionary leap that dramatically increases the feasibility of on-device AI. The ability to leverage advanced AI securely without data ever leaving the device offers immense benefits. This innovation resolves cloud API costs and data transfer issues associated with agent AI running multiple tasks, opening doors for SMBs to utilize more sophisticated AI cost-effectively and securely. 📎 Read more
4️⃣ Long-Horizon-Terminal-Bench: New AI Agent Evaluation Standard
🏷️ AI Models, Large Language Models What happened? Researchers from Tencent Hunyuan introduced "Long-Horizon-Terminal-Bench," a new benchmark to measure the true capabilities of AI agents in long-duration tasks. Comprising 46 complex tasks across experiment reproduction, software engineering, and scientific computing, it uses a "dense reward system" to evaluate intermediate progress, not just final outcomes. Even frontier models achieved only a 15.2% pass rate, indicating significant room for improvement in long-horizon AI agent tasks. Our take While previous AI evaluations often focused on short, simple tasks, this new benchmark rigorously assesses AI agents' capabilities in more realistic, complex, and time-consuming scenarios. This will encourage AI agent developers to focus not just on improving benchmark scores, but on enhancing truly practical abilities like long-term planning and context management, which are crucial for real-world applications. 📎 Read more
5️⃣ Meta Releases DINOv3, Revolutionizing Image Recognition with Self-Supervised Learning
🏷️ AI Models, Multimodal AI What happened? Meta open-sourced DINOv3, a state-of-the-art general-purpose computer vision model trained with self-supervised learning (SSL), under a commercial license. Trained on 1.7 billion images with 7 billion parameters, a single backbone outperforms specialized solutions in high-density prediction tasks like object detection and semantic segmentation. By eliminating the need for data labeling, it has the potential to significantly reduce AI adoption costs and time. Our take DINOv3 could be a game-changer for image recognition AI, substantially lowering the barrier to entry. For SMBs, in particular, which often struggle with acquiring high-quality labeled data, SSL-based training opens new avenues for AI utilization. Efficient and accurate AI deployment is expected across all operations where visual information is critical, such as product quality inspection, anomaly detection in factories, and customer behavior analysis in retail. 📎 Read more
6️⃣ NVIDIA Nemotron Gains Traction in Japanese AI Ecosystem
🏷️ AI Models, AI Infrastructure, Asian AI ecosystem, NVIDIA What happened? NVIDIA's open AI model, Nemotron, is being adopted by major Japanese companies and startups including Sakana AI, SoftBank, NTT, and NEC. Sakana AI, in particular, is integrating Nemotron into "Fugu," its unique layer for dynamically routing models. Nemotron is gaining attention as a foundational platform that enables companies to build AI models and applications tailored to their specific language and industrial characteristics, accelerating AI development in Japan. Our take The expansion of NVIDIA Nemotron in the Japanese market highlights the importance of AI development tailored to Japan's unique language and industrial needs. By leveraging open models, companies can customize and manage AI in-house, making it easier to deploy localized AI solutions for challenges like an aging population and labor shifts. The combination with orchestration layers like Sakana AI's Fugu will foster a mature Japanese AI ecosystem by enabling flexible AI utilization that is not dependent on a single model. 📎 Read more
7️⃣ MCP Protocol to Enhance Scalability with 2026 Revision
🏷️ AI Infrastructure What happened? The Model Context Protocol (MCP), a common language for connecting AI models with external tools, is undergoing a major revision by July 2026. The biggest change is a shift from the current stateful, session-dependent connections to a stateless model, allowing MCP servers to scale horizontally behind load balancers. This will resolve scalability issues for MCP in large-scale AI systems, enabling more flexible operations. Our take The evolution of MCP, the foundation for AI integration with existing systems, will significantly impact SMBs' AI adoption strategies. Improved scalability through statelessness will directly translate to reduced operational load and improved cost-efficiency for companies considering cloud-based AI. If you plan to embed AI deeper into your operations, it's worth monitoring this protocol's developments and assessing your systems' compatibility with the new MCP. 📎 Read more
8️⃣ Apple Sues OpenAI Over Trade Secret Theft Allegations
🏷️ Big 3 coverage, controversy What happened? Apple has filed a federal lawsuit against OpenAI, alleging trade secret theft. Apple claims OpenAI stole its intellectual property and used it to develop its own consumer hardware. While the two companies previously collaborated on integrating ChatGPT into iPhone OS, their relationship reportedly soured after OpenAI's plans to enter the hardware business. The complaint also alleges OpenAI instructed former Apple employees to share Apple's confidential information during interviews. Our take This lawsuit between former collaborators Apple and OpenAI symbolizes the intensifying competition and intellectual property protection concerns within the AI industry. OpenAI's plans to enter the hardware business appear to be a key factor in the deteriorating relationship. This case is a significant example of how partnerships and competition among companies are shifting with AI's rapid advancement, potentially having a major impact on the future direction of the AI industry. 📎 Read more
9️⃣ XPENG Accelerates "Physical AI" Strategy, Unveils Humanoid Robot IRON
🏷️ Asian AI ecosystem What happened? Chinese EV giant XPENG declared its transformation into a "Physical AI" company at its "XPENG LIVE: PHYSICAL AI FOR ALL" event. It unveiled its next-generation humanoid robot, IRON, and the globally launching autonomous SUV, MONA L03. The core of its Physical AI strategy, aiming for mass production of humanoid robots by late 2026, is the "Vision-Implicit Token-Action (VLA 2.0)" model, which generates action instructions directly from visual signals. The livestream garnered over 3.4 million views in four days. Our take XPENG's "Physical AI" strategy clearly demonstrates AI's evolution towards applications involving physical actions in the real world. The humanoid robot IRON and autonomous vehicle MONA L03 could significantly impact operational efficiency across various industries, including manufacturing, logistics, and services. Direct control technologies like VLA 2.0, which translate "vision to action," will be essential for AI to handle physical tasks in the future. SMB leaders should start considering how such technologies could be applied to their physical operations now. 📎 Read more
🔟 Google's Gemini 3.5 Pro Delayed Due to Code Generation Shortfalls
🏷️ Big 3 coverage, controversy, Large Language Models What happened? Google's flagship AI model, Gemini 3.5 Pro, has reportedly been delayed by several months due to insufficient code generation capabilities. According to Bloomberg, Gemini 3.5 Pro has not met internal targets and lags behind recently released AI models from competitors like OpenAI and Meta in software code generation. This news led to a 4% drop in Alphabet's stock price. Our take The delay of Google Gemini 3.5 Pro highlights the intensity of AI development competition and the critical importance of code generation capabilities for frontier models. As OpenAI and Meta make rapid advancements in this area, Google's challenges may signify more than just a technical delay. Since an AI model's capabilities directly translate to its practical utility and market competitiveness, attention will be on Google's future moves and how it plans to close this gap. 📎 Read more
📚 Editor's Note
This week in AI clearly showed a trend where open-source models are approaching the performance of frontier models and expanding their practical application range. The emergence of models like Moonshot AI's Kimi K3, Thinking Machines' Inkling, and the iPhone-operable Bonsai 27B signals the arrival of an era where AI is democratized, and more developers and businesses can access cutting-edge AI technology.
Meanwhile, the evolution of AI agent evaluation standards and the revision of the MCP protocol indicate that AI technology is shifting its focus beyond mere model performance to operational usability, scalability, and real-world problem-solving capabilities. News such as the Apple-OpenAI lawsuit and the delay of Google Gemini 3.5 Pro underscored the fierce competition, intellectual property protection concerns, and technical challenges underlying this rapid evolution.
In the realm of physical AI, XPENG's unveiling of the humanoid robot IRON suggests that a future where AI performs physical actions in the real world is steadily approaching. These developments foreshadow that AI will be deeply integrated into our work and lives at a pace faster than we might imagine.
Until next week, we'll continue to bring you useful insights and thought-provoking perspectives. dera news editorial team
📬 About this newsletter