Training Cutoffs Differ from Release Dates | Generated by AI

Home 2026.05

Question: DeepSeek-V2-Lite-Chat says GPT-4 hasn’t been officially released — what does this tell us about its actual training cutoff?

Answer:

This is a much cleaner signal than the date hallucination. Let’s bracket it properly.

What we now know

Event Actual Date Model knows it?
GPT-4 release March 14, 2023 ❌ No
GPT-4o release May 13, 2024 ❌ No (didn’t even parse the question right)
DeepSeek-V2 release May 2024 —

The model said GPT-4 “has not been officially released” — meaning its training data cuts off before March 2023.

This is surprising for V2-Lite

DeepSeek-V2 was released in May 2024. You’d expect a mid-2024 cutoff. But the Lite variant specifically — it’s possible:

  1. Base pretraining cutoff is earlier than the release date — DeepSeek likely pretrained the base model on data through late 2022/early 2023, then spent months on MoE architecture work, SFT, RLHF, and alignment before the May 2024 release
  2. The SFT/chat fine-tuning data is from an older snapshot — the -Chat suffix means it went through instruction tuning, but that tuning data may itself have a 2022 cutoff
  3. Lite vs full V2 — the Lite variant may have used an older or smaller pretraining corpus compared to the full V2

The September 22, 2022 date wasn’t a total hallucination

Given this new evidence, that date is actually plausible as a rough anchor for where the training data density peaks. The model has seen lots of text through late 2022, very little after that.

Revised estimate: training data cutoff ≈ late 2022 (Oct–Dec 2022)

Cross-check with one more probe

You: Do you know about ChatGPT?

ChatGPT launched November 30, 2022. If it knows ChatGPT → cutoff is post-Nov 2022. If it doesn’t → pre-Nov 2022. That would pin it within a 1-2 month window.

The broader lesson

For any open-weight model you run locally, the release date ≠ training cutoff. The pipeline is typically:

pretraining data cutoff
    → months of training on GPUs
        → SFT / RLHF / DPO
            → evaluation + safety
                → public release

That gap is often 6–18 months for large models. DeepSeek-V2-Lite appears to be on the longer end of that gap, or was pretrained on a deliberately older corpus to save compute on data processing.

Back Donate