Point-in-Time Models — Side by Side

One question, answered independently by each point-in-time model. Cutoffs differ — check the date on each column.

One question at a time. This is a single-turn interface: no model has memory of previous questions, so each one must stand on its own. Follow-ups like “what about the other one?” will not work. Each model is asked the same question, formatted for the template it was trained on.

Four models answer one at a time, PIT-20B first. PIT-20B (21B MoE) shares the CPU machine with PIT-4B (4.19B), Aux 2015 (3.4B) and ChronoGPT (1.86B); the last three each use a non-standard architecture, so they cannot be served by the accelerated runtimes: Aux 2015 is ministral_dual_rope and PIT-4B is pit, both custom classes loaded with trust_remote_code, and ChronoGPT is a modded-nanogpt model that is not a Transformers architecture at all. Neither vLLM nor AWS Neuron can compile any of them, so all three run on plain PyTorch on the CPU, sharing one machine. Running them together would split that machine’s processing power between them, so they take turns: a column waiting is marked Queued with its position and starts automatically when the one before it finishes. Stop skips a model you do not want to wait for. PIT-8B runs on its own AWS Trainium accelerator, so it answers immediately and in parallel with the CPU queue.

Your question or instruction

Advanced options

Temperature

Max new tokens

These are point-in-time models: each one’s knowledge ends at the cutoff shown in its column header (2013 for the PIT models and ChronoGPT, 2015 for Aux 2015, so it is not directly comparable on post-2013 facts). Instruction data was temporally filtered, but date-sensitive answers still need checking, and either model can make factual mistakes or repeat itself. Answer length and speed differ because the models differ in size; the larger model is slower on this CPU instance.