DeepSeek Just Made Its Own Flagship Model Obsolete

DeepSeek dropped a new model today, and the real story here is not some benchmark chart. It is how much cheaper and faster they made it while still climbing that chart.

V4.1 Flash is the small, fast member of DeepSeek’s new model family, and it can now actually look at images, not just read text. Under the hood it is a mixture-of-experts setup with more than five hundred billion parameters total, but only a handful light up for any single answer. That is the trick that keeps the cost down while the model still holds a full million tokens of context when you need it.

Here is the part that actually matters if you build on this stuff. DeepSeek says internal testing shows the new small model beats their own bigger V4 Pro on speed, cost, and quality all at once. Starting in a few days, if you send a request to V4 Pro, DeepSeek is just going to quietly route it to this smaller model instead and charge the smaller model’s price. Their own flagship goes obsolete inside their own lineup within days of a cheaper model landing.

And they open sourced the whole thing. Full weights are up on Hugging Face under an MIT license, so anyone can download it, run it locally, fine tune it, build a business on top of it, whatever they want.

This is what real competition in AI looks like day to day. Not a keynote once a year, just a lab shipping a smaller, cheaper, better model on a random Wednesday and giving it away for free while they are at it. Prices for everybody building on these models keep sliding down because of exactly this kind of release, and that is good news if you are the one paying the API bill.

——

Follow: @Ali Demi
Book your free AI clarity call, NOW!
https://buff.ly/TpWy277

——

Sources:
https://www.deepseek.com/en/news/deepseek-v4-1-flash/
https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
https://www.marktechpost.com/2026/09/10/deepseek-ai-released-deepseek-v4-1-flash-with-1m-context-fp4-kv-cache-and-cross-layer-attention-reuse/

Repost this. Thanks.