AI developer tools · What shipped

oMLX 0.7.0 released with significant performance improvements across multiple AI models including faster Qwen3.8, GLM-5.3-Flash and MiMo V2 inference. The update features a completely rebuilt memory guard system, improved prefill and decode operations with optimized kernel fusions, and support for Lightning MTP with expert offload for faster model generation. Key performance gains include 31.9-79.0% faster prefill speeds and substantial decoding improvements across different model architectures and context lengths.

Apple released Core AI Models, an open source repository providing model export recipes, Python primitives, and Swift runtime utilities for building on-device AI. The project includes a curated model catalog with export recipes from Hugging Face, reusable PyTorch building blocks for authoring custom models, Swift package utilities for macOS and iOS integration, and agent skills plugins for coding agents like Claude Code, Codex CLI, and Gemini CLI. Supported models are exported as standalone .aimodel files with accompanying resources, and CLI tools are provided for running exported models directly on Mac with Xcode 27.0+.

Google announced Gemini 4 Argon, a frontier AI model positioned for enterprise workflows including software engineering, cybersecurity, and business automation. Across 18 disclosed benchmarks, Argon leads or ties in 13 categories, outright leading on 12—posting particularly strong results on DeepSWE v1.1 (77.9% for long-horizon coding tasks), AutomationBench (51.3% for business execution), and Harvey's Legal Agent Benchmark (19.6%). The model supports 1 million output tokens and will launch at $2 per million input tokens and $10 per million output tokens during an introductory pricing period, dropping to $4/$20 after. However, Argon is beginning with limited rollout to trusted cyber defenders through Google's Fairwind Program and U.S. government pre-release access, with broader API availability planned for paid customers and Google AI Ultra subscribers. Compared to GPT-6 Astra and Claude Opus 5.5, Argon shows strongest margins in enterprise and long-context workflows but trails on certain software and science benchmarks, indicating workload-dependent model choice remains necessary for developers and enterprises.

Kong Inc. announced Volcano, an AI-native developer platform that provides unified infrastructure for building, deploying, and operating AI agents and modern web applications. Volcano integrates durable compute, branchable PostgreSQL databases with vector support, edge functions, frontend hosting, authentication, real-time services, file storage, and coordination tools like locks and queues in a single platform. The platform ships with first-class integrations for Claude Code, OpenAI Codex, and Cursor, allowing developers to deploy agents to production without assembling infrastructure from multiple providers. Key capabilities include durable workflows with built-in leader election, autoscaling edge functions across regions, global CDN-backed file storage, and multi-agent coordination features.