DeepSeek's V4-Flash-Vision-Exp repo is code, not weights

DeepSeek put a new repo on Hugging Face, and it reads like a model drop without being one. The Hugging Face repo has a tokenizer, a prompt-encoding reference and a minimal PyTorch implementation covering the vision encoder and aligner, DFlash attention, MoE routing, Hyper-Connections and the DSpark forward path. No weight files.

That distinction matters because V4-Flash-Vision-Exp is a 284B-parameter mixture of twenty 13B experts. Nobody is running that locally off a code reference. The model has been API-only since August 21, priced at $0.14 per 1M cache-miss input tokens and $0.28 per 1M output tokens, with images billed at up to 384 tokens each and no separate per-image fee.

A thread about the repo picked up 228 upvotes with people treating it as an open-weights release. It isn't. What you get is enough to read how the architecture works, port pieces of it, or write a compatible client, not enough to load the model yourself.

This is a pattern worth knowing if you track model releases for what you can actually self-host: a code-only reference repo next to a paid API is not the same commitment as an open-weights drop, even when the file browser looks similar and the excitement in the comments doesn't distinguish between them. If you're evaluating whether a "release" changes your deployment options, check for weight files before you check the upvote count.