Tuesday, September 1, 2026

Flash Next MTP no longer dumps state every round. Metal 4.0 tensors light up on M5.

llama.cpp unblocked Flash Next MTP speculative decode — it no longer dumps recurrent state to host every round. Speed numbers are CUDA contributor benches, not Mac; community MTP GGUF, not a first-party path. Same morning, Metal 4.0 tensor API actually compiles on M5 and A19 so prefill can use Neural Accelerators; has_tensor had been silently off on every device. Does not help M3 Ultra. Try-this is M5/A19 only. Elsewhere: AWS Agent Registry GA (Quick can discover those records), BigQuery Graph GA, Databricks Unity Catalog Skills Beta (Issue 7 miss). From the field: Experiential (Apache-2.0) and tare (MIT). NVIDIA/HF still unconfirmed. No frontier flagship. No new OSS tools or models GA.

Window ~2026-08-01 to 2026-09-01. Monday already covered Ultra Metal tables, Flash Next n-gram, Ollama 0.33.2, Hy4, Vision-Exp, Infer-forge, AgentCore Memory FGAC, Genie One, EU DSA ChatGPT. Primary sources only. Vendor evals flagged.