AI Daily · 2026-07-28

Moonshot AI released Kimi K3, a 2.8-trillion-parameter native multimodal model with novel KDA and AttnRes architectures and a 1M-token context window, matching top closed models like Claude Fable 5 and GPT-5.6 Sol on long-form coding, agentic knowledge work, and multimodal benchmarks. The weights are shared under an “open weight” license that requires commercial entities with sufficient revenue to negotiate separately, while third-party inference is already live at the same pricing. Other updates include Cursor’s ₹649 Start plan for India, and LangChain’s shift to an agent-first data stack along with a new full-text search design for SmithDB.

East Asia · First-hand

Moonshot AI

⭐⭐⭐⭐⭐ [Model Release] moonshotai/Kimi-K3

Moonshot AI released Kimi K3, its most capable model to date: an open-weight 2.8T-parameter native multimodal agentic model built on the new Kimi Delta Attention (KDA) and Attention Residuals (AttnRes) architecture, with a 1M-token context window and native image–text understanding. The model excels at long-horizon coding, agentic knowledge work, and multimodal tasks, with benchmark results rivaling leading models such as Claude Fable 5 and GPT-5.6 Sol. The full weights are openly released under the Kimi K3 license to foster research and innovation.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐⭐ [Model Release] moonshotai/Kimi-K3

Moonshot released the weights for Kimi K3, a 2.8 trillion parameter model, with the 1.56TB files now available on Hugging Face. The license no longer claims to be “modified MIT” and now requires a separate agreement with Moonshot for large Model-as-a-Service businesses that exceed revenue thresholds; Moonshot consistently describes it as “open weight” rather than open source. OpenRouter already offers K3 from 7 providers at the same $3/million input and $15/million output pricing as Moonshot itself.

Product Update

⭐⭐ [Product Update] Cursor Start

Cursor launches the 'Start' plan for developers in India at ₹649/month with local UPI payment support. It includes generous access to Grok 4.5 and Composer models, always-on cloud agents, iOS remote control, and extensibility via plugins, MCP servers, hooks, and skills. Existing free users can upgrade from the dashboard, and new users can select it during signup. The plan is available from July 28, 2026 with monthly auto-renewal.

⭐⭐ [Product Update] How LangChain Built an Agent-First Data Stack

LangChain's data team migrated from a traditional BI tool to an agent-first self-serve data stack, equipping the agent with clear data models, metric definitions, business context, and trusted sources for more accurate answers. The agent now handles approximately 40x the request volume that the three-person data team could field directly, with nearly 100% of provisioned users (a third of the company) engaging across ~2,200 conversations in 30 days. The data team's role has shifted from answering every question to continuously improving the system through modeling, guardrails, and feedback loops. This architectural shift demonstrates how injecting rich context helps agents deliver more trustworthy and broad self-service analytics.

⭐⭐ [Product Update] Full Text Search in SmithDB: Designing an Inverted Index for Object Storage

LangChain's SmithDB now supports full-text search and JSON filtering over agent traces, delivering a median (P50) latency of 400 ms even on large, deeply nested JSON documents stored in object storage. The post explains the design of an inverted index tailored for object storage, balancing flexibility with query speed. This implementation detail is useful for developers working within the LangChain ecosystem.

Opinion

⭐⭐ [Opinion] Own Your Intelligence: The Key to Lasting AI Advantage

Generic AI alone cannot create lasting advantage; companies must own their intelligence by controlling models, agent systems, context, and memory. This involves managing cost, quality, risk, and behavior while establishing a feedback loop that compounds intelligence with use. Using examples like insurance claims and vertical AI startups, the article argues that differentiation comes from tailored intelligence rather than the same base model APIs, and companies should build their own intelligent systems on purchased infrastructure.

⭐⭐ [Opinion] An opinionated guide to which AI to use to do stuff

Simon Willison comments on the evolution of Ethan Mollick's AI usage guide: a year ago it centered on chat models, now it focuses on agentic systems that can autonomously complete hours of work. Google Gemini dropped off the list due to lack of a mature agent offering, with Gemini Spark yet to prove itself. The piece dissects the confusing naming of ChatGPT and Claude modes like 'Work', 'Cowork', and 'Codex', highlighting how the same mode names behave differently between mobile and desktop, such as mobile ChatGPT Work actually enabling internet access for Code Interpreter.