https://radar.montevive.ai/ What actually moved in LLMs, once a month. llama.cpp, small-model compression, and the open weights worth downloading. 2026-08-16T12:00:00Z Montevive https://montevive.ai ai@montevive.ai © 2026 Montevive https://radar.montevive.ai/issues/2026-08-16-llm-state-of-the-art.html 2026-08-16T12:00:00Z 2026-08-16T12:00:00Z

Covering 16 July to 16 August 2026. llama.cpp got NVFP4 end to end plus a KV cache cloning endpoint reporting up to 2.12x on shared-prefix workloads, and suffix decoding brought speculative decoding without a draft model. On-policy distillation displaced structured pruning as the way to make a small model good, and KV cache compression is where the research volume went. Four independent papers found compressed models passing every standard quality guard while behaving measurably differently. Includes a glossary of the underlying technology.