Community
DeepSeek V4.1 Flash 1017 identifies three of four small RL tasks from raw weights
The model was given four anonymized 51k-parameter RL checkpoints as text and correctly identified CartPole, Flappy and Catch but confused BallBeam with CartPole. The tester calls it preliminary, with one run per checkpoint.
Every outlet that ran it 1 article
r/LocalLLaMACommunityLead
DeepSeek V4.1 Flash 1017 can also do Weight-Space Programming