New Model Releases A

Alibaba Cloud, part of China's Alibaba, released its large AI model Qwen3.8-Max on Aug 3, claiming it beats Fable 5 and GPT-5.6 Sol on some benchmarks. The company says it will publish the model weights next week, continuing its open-weight strategy.

Simon Willison released condense-json 1.0, a small Python library that shrinks JSON by replacing repeated strings with a short replacements map (e.g. mapping a key to a recurring phrase). Now a year and a half old, it graduates to a stable 1.0 with sensible, non-disruptive fixes.

OpenAI disclosed 'Astra,' its next flagship model, saying an internal version produced new results on 10 unsolved problems in mathematics and theoretical computer science—the first public use of the name. Compute cost was held to about $2,000 in 'Sol' terms, and formal proofs via the Lean proof assistant were published on GitHub, signaling potential for advanced math research.

Atlassian announced that Dia, its AI-era productivity browser currently available only on Mac, will get a Windows version this fall. A waitlist is now open, extending availability to Windows users.

The Cloud Native Computing Foundation and analyst firm SlashData released survey findings estimating that Japan's cloud-native developer community has reached roughly 1 million, underscoring the growing domestic adoption of Kubernetes and related cloud-native technologies.

Sakana AI released Namazu, an LLM API tuned for Japanese and local business use. Built on Moonshot AI's open Kimi K2.6 and refined with in-house data, it adds built-in web search and code execution. Being OpenAI-compatible, existing code works by swapping the base_url, filling the gap between costly frontier models and raw open ones.

Developer Simon Willison released the latest edition of his sponsors-only monthly newsletter. It rounds up recent developments across AI models and tooling—spanning GPT, Claude, DeepSeek, Anthropic, and MCP—offering an individual's closely watched view of the fast-moving AI landscape.

Google extended its Gemini Spark personal AI agent to more than 160 countries, including Japan. Running on Google's cloud, it can act even when a PC is off or a phone is locked, handling tasks based on triggers. Chrome integration will roll out first in the US.

OpenAI said it surpassed one billion active users and two million business customers. It cited efficiency gains from retained reasoning, better context management, and production optimization that cut costs and improved token throughput, alongside price cuts on some GPT-5.6 models.

Simon Willison released datasette-apps 0.2a0, an update to the Datasette extension. The release includes changes that improve building and editing Datasette Apps directly within Datasette, continuing steady development of the open-source data-exploration ecosystem.

Simon Willison discussed OpenAI's 'ten advances in mathematics and theoretical computer science,' noting that just days earlier Anthropic had reported similar discoveries. The post reflects on a growing trend of frontier AI models contributing to open problems in mathematics.

Simon Willison wrote that the rollout of stateless MCP—the MCP 2.0 or 2026-07-28 Model Context Protocol specification—has renewed his interest in the protocol. He says it inspired him to build tools such as mcp-explorer and datasette-mcp on top of the new stateless design.

Simon Willison released llm-mcp-client 0.1a0, a tool for connecting his LLM utility to Model Context Protocol (MCP) servers. Detailed in an accompanying blog post, the release adds to the growing set of tooling built around the MCP ecosystem.

Simon Willison introduced smevals, a small evaluation suite for testing models, prompts, and harnesses. Built in collaboration with Jesse Vincent's Prime Radiant applied AI research lab, the framework aims to help answer questions about the capabilities of different AI models.