Tuesday, August 25, 2026
llama.cpp cut v0.3.0 this morning with Metal FlashAttention vec numbers for M1 Pro, M2 Ultra, and M5 Max. MLX 0.32.2 is on PyPI. mlx-lm merged Muse Glimmer and Ling-3.0 on main (PyPI still 0.31.3). AWS published ARD, an open spec for federating agent catalogs. Late Monday: ADK live/voice eval, Groq 3 LPX production, HyperPod Ray, GPT-5.6 in Kiro and GovCloud, Agent Lightning’s first Skill, Thomson-1.0-Small.
Window ~2026-07-29 to 2026-08-25. Monday already covered Claude computer use, llama.cpp GLM-4.5-Air MTP, and SenseNova on MLX. Primary sources only. Vendor evals flagged.
Breaking AI News
Published 2026-08-25 (~6:22 AM ET); matching nightly is b10621. Changelog includes dots3-note, GLM-4.5-Air MTP (already in Monday’s b10603), DeepSeek 4 -sm tensor, and ggml v0.22.0 with per-op Metal kernels. PR #26570 is the first maintainer end-to-end Metal FA-vec table: M2 Ultra qwen3 0.6B Q8_0 tg32 ~1.3–1.5× at 2K–32K vs master; no gain at d=0. Still no Metal MTP numbers. vendor
24 Aug: Apache 2.0 spec for describing and discovering agents, MCP servers, tools, and skills. Normative text at agenticresourcediscovery.org and ards-project/ard-spec. AWS frames it next to AWS Agent Registry (preview since 9 Apr 2026); do not treat Registry as an Aug 24 launch.
