Kimi K3 Launches with 1 Million Context Tokens, Escalating the AI Arms Race That Moves Crypto Markets Moonshot AI has unveiled its latest AI model, Kimi K3, which boasts a context window of up to 1 million tokens. This represents a fourfold increase compared to its predecessor, the K2 series, which supported context windows of approximately 256,000 to 262,000 tokens. The new model is now available across web platforms, apps, and APIs, marking a significant step in the company’s rapid development. The core innovation of Kimi K3 lies in its expanded context window, which allows the model to process vast amounts of information simultaneously. This capability is particularly useful for tasks requiring the analysis of large codebases or entire books. The model is expected to operate on a Mixture-of-Experts (MoE) architecture, utilizing approximately 2 to 3 trillion parameters. This is a substantial jump from the K2 series, which ran on about 1 trillion parameters. The MoE approach routes queries to specialized sub-networks, enhancing efficiency by not activating all parameters for every task. Another standout feature of Kimi K3 is its Agent Swarm technology, which enables the model to coordinate up to 300 sub-agents working in parallel. This allows for complex multi-step planning and execution, positioning the model as a powerful tool for advanced applications. Early access to the model reportedly began rolling out around July 14 to 15, with Moonshot AI making it available through its web platform, desktop CLI, and API endpoints. Moonshot AI, founded in March 2023 in China, has demonstrated rapid growth. By October 2023, the company had launched its Kimi chatbot, initially supporting 128,000 tokens. The K2 series was introduced in July 2025, followed by iterative improvements. K2.#hugging_face #moonshot_ai #kimi_k3 #anthropic_claude_opus #moonshot_ai_china
Google's New Gemma 4 Models Bring Complex Reasoning Skills to Low-Power Devices Google LLC has launched its latest open-weight artificial intelligence models, Gemma 4, marking a significant advancement in the field of lightweight, high-performance AI. These models, built on the architectural foundation of Gemini 3, are designed to handle complex reasoning tasks and support autonomous AI agents running on low-power devices such as workstations and smartphones. The release positions Google as a key player in the growing market for edge computing and local AI applications. The Gemma 4 family includes four variants: Effective 2B, Effective 4B, a 26B Mixture of Experts (MoE) model, and a 31B Dense model. The smaller "Effective" models are tailored for edge use cases, such as Android smartphones and Raspberry Pi computers, while the 26B MoE model introduces an innovative approach by activating only 3.8 billion parameters during inference tasks. This optimization allows the model to maintain high performance without compromising the depth of knowledge typical of larger models. The 31B Dense variant currently ranks third in open models on the industry-standard Arena AI Text leaderboard, demonstrating its competitive edge. Google DeepMind researchers Clement Farabet and Olivier Lacombe highlighted the models' ability to deliver "more intelligence per parameter," enabling them to outperform their size class. This efficiency is critical for applications requiring real-time processing and minimal computational resources. The models are also engineered to support AI agents, with native capabilities for function calling and structured JavaScript Object Notation (JSON) outputs.#google_llc #google_deepmind #clement_farabet #olivier_lacombe #hugging_face
