Tag
A user reports that the latest LM Studio update (0.4.17) eliminated the multi-token prediction speed boost, reverting to previous performance on an RTX 5090 setup.
The author trained 1B, 2B, and 3B models with the same SFT recipe and observed that instruction-following (IFEval) regressed for the 1B and 2B models but improved for the 3B, possibly due to different learning rates or model capacity.
Claude Opus 4.7 shows decreased performance compared to versions 4.6 and 4.5 on SimpleBench evaluation.