KVCache.ai launches open-source web-based KV Cache Size Calculator for models including DeepSeek V4 Flash, Qwen3, GLM, Kimi, and MiniMax

I think Whale will release a V4 model with 10M context this year The economics make sense now

Nearly a 100x difference in KV cache size of Minimax 2.7 & V4 Flash

When DeepSeek launches their harness, they probably launch a plan too. I think it would be neat for them to offer a plan heavily leveraging their unique economics. For example, a growing user memory up to 2M tokens. No need for special pricing: nobody else can do that anyway.

@zephyr_z9 I don't think so. I bet they'll converge to the same learned /compaction feat as Composer/Codex/Claude have done.

This makes the deployment both cheaper and less compute-intensive

@teortaxesTex I have been dreaming about some sane calc finally, but why there are so few models to choose from?

@zephyr_z9 Also comparing BF16 vs FP8/4 mixed on a sliding window.

Deepseek definitely cooked, but it's not apples-to-apples.

@zephyr_z9 Wouldn't it be correct to compare them while keeping the precision?

Even so, there is a significant difference anyway. 🥲

@AgiWhen @teortaxesTex I thought most open-weights labs do

@vacla_vcerny @teortaxesTex maybe because most labs do not disclose the vector dimension and layers, so it's hard to predict.

@teortaxesTex tbh i didnt know u could calculate that granularly

the deepseek V4 sparse attention detail is what makes it interesting

@teortaxesTex It would make sense if they can get the data

@teortaxesTex DeepSeek's cache pricing finally surfaces what V4-Pro did under the hood.

10% of V3.2's KV cache at 1M tokens. Long context got cheap through smarter cache, not bigger HBM.

@zephyr_z9 Isn't that indicates accuracy and performance might drop? Especially if we get rid of boring tokens, the whole context wouldn't be preserved. I'm sure they know the downside and tryna minimize it but hopefully new breakthrough appears!

@teortaxesTex i didnt like when they said they want to hire someone to code a harness like claude code, they should've said codex

@zephyr_z9 We don't have training data for maintaining coherence at 10M context

@zephyr_z9 Less than a year ago I was barely hopeful for 256k context. 🥹

@teortaxesTex Will closed frontier model do the same tricks with KV cache?

@zephyr_z9 100x tracks. DeepSeek V4-Flash runs 7% of V3.2's KV cache at 10% FLOPs. Smarter cache, same context window.

@zephyr_z9 Moar nand

@zephyr_z9 they are planning to release a competitor to claude code/codex this year presumably

so they need a coding focused model like v4 series without the need for 10m context, though more capable than existing version (despite being token efficient) https://the-decoder.com/deepseek-wants-to-take-on-claude-code-and-openais-codex-with-deepseek-code/