AI developer tools · What shipped

llama.cpp released build b10603 with infrastructure improvements including an optimization to skip device info loops when not needed for logging, avoiding unnecessary GPU context creation and 550 MB VRAM allocation with CUDA backends. Recent builds added PAD_REFLECT_1D operation support for Vulkan backend with performance testing on Intel Iris Xe showing 5.38-158.31 microseconds per run. The project maintains multi-platform support across macOS, Linux, Android, Windows, and openEuler with various backend options including CUDA, Vulkan, ROCm, OpenVINO, and SYCL.

Google announced Gemini 3.7 Flash, a new AI model optimized for coding and agent tasks across Google AI Studio, Google Antigravity, and the Gemini API. The model demonstrates significant performance improvements on software engineering benchmarks: 65.3% on DeepSWE v1.1 (versus 49.0% for 3.6 Flash) and 43.6% on FrontierCode 1.1 Main (versus 34.4% for 3.6 Flash). Key developer features include better code generation with autonomous error recovery, improved design adherence in web development (Elo score of 1588 on WebDev Arena), and configurable thinking levels (low, medium, high) for latency versus accuracy tradeoffs. Pricing starts at $0.75/1M input tokens and $3.75/1M output tokens through December 31, 2026. The model is available today in Google AI Studio, Google Antigravity, and through integrations with frameworks like LangChain and LiteLLM.

Vercel released Is Agentic, a free tool that scores how readily AI agents can discover, access, understand, and use public websites using 118 checks across four weighted layers: Discovery (20 points, 15 checks), Access (30 points, 41 checks), Usability (40 points, 56 checks), and Payments (10 points, 6 checks). The methodology was reverse-engineered from real agent runs by Ora, an agent-experience research company. The tool is available at zero cost through a browser interface, CLI (npx is-agentic), read-only API at /api/v1/report (120 requests per IP per 60 seconds), MCP server, and official agent skill, with reports delivered as HTML, Markdown, JSON, and CLI output. Vercel regroups Ora's checks into Essential (80 points), Recommended (20 points), and Emerging signals (bonus up to 5 points), excluding not-applicable checks so sites are not penalized for missing interfaces they never claimed to offer.