Monday, August 31, 2026

M3 Ultra gets its own Metal fa-vec tables. Flash Next long-context gen got cheaper.

Friday said an M3 Ultra does not inherit M3 Max Metal tables. Sunday it got its own. llama.cpp also sped up Flash Next generation as context grows — contributor CUDA/ROCm benches only, not Mac numbers. Ollama v0.33.2 is now stable. Try-this is the same Unsloth Flash Next GGUF on a 256 GB Mac, now with Ultra tables. Build numbers and types are in Local and MLX. Elsewhere: Hy4 preview Apache-2.0, DeepSeek Vision-Exp MIT weights on Hub today, LMSYS Infer-forge methodology (not OSS, not a new SGLang tag), AgentCore Memory access control, Databricks Genie One, and the EU DSA designation of ChatGPT. From the field: Terminal-Bench-Science, Headlong, and OneCLI. NVIDIA/HF still unconfirmed. No frontier flagship.

Window ~2026-08-01 to 2026-08-31. Friday already covered Flash Next native + Unsloth GGUFs, Max/M5 Metal fa-vec, MiniMax-H3 on H200s, GLM-5.3-Flash MIT. Primary sources only. Vendor evals flagged.