Nemotron IMO Gold: How NVIDIA Trained an Open Model for Olympiad Proofs
How the Nemotron IMO system reached 30 of 42: SFT, RL and a generate-verify-refine loop, with the numbers worked through and caveats.
Open models
Kimi K3.1 Pro Weights Are Public. The License Decides Who Can Use Them
Kimi K3.1 Pro weights are public, but the license sets revenue and user thresholds. What the terms say, how the benchmarks hold up, and what…
DeepSeek V4.1-Flash-1017: What the October Update Changed in the Open-Weight Model
DeepSeek’s Oct 17 V4.1-Flash-1017 update keeps MIT weights and an 890-byte KV cache per token while lifting agentic benchmark scores. Specs and run notes.