All Chinese AI models are now switching to a combination of KV caching and sparse attention blocks. Huge speedups. This is https://www.minimax.io/
DeepSeek 4 is also doing something like this. 1 billion tokens is now just 15$
https://social.here.blue/media/7da6d2fc58670736c66f0c8bf571cbb7ee28c9399dd8df010d6376bb5a201181.blob
Theodore John Kaczynski
npub1eg…k9zfr
2026-05-27 00:37:48 UTC
Author Public Key
npub1egvswnt7kelt7x9wew3sakuavpq8a4kg9leyp59xqfj5rwnayq2qgk9zfrSeen on
wss://relay.ditto.pubPublished at
2026-05-27 00:37:48 UTCEvent JSON
{
"id": "7418b14d13fbd2571a52564953a5acfc92140ef578d1cf6deb840ec7af09a338",
"pubkey": "ca19074d7eb67ebf18aecba30edb9d60407ed6c82ff240d0a6026541ba7d2014",
"created_at": 1779842268,
"kind": 1,
"tags": [
[
"imeta",
"url https://social.here.blue/media/7da6d2fc58670736c66f0c8bf571cbb7ee28c9399dd8df010d6376bb5a201181.blob",
"m image/png"
],
[
"proxy",
"https://social.here.blue/objects/8cbbf531-cfac-4758-beb0-99a0e3af47f6",
"activitypub"
],
[
"client",
"Mostr",
"31990:6be38f8c63df7dbf84db7ec4a6e6fbbd8d19dca3b980efad18585c46f04b26f9:mostr",
"wss://relay.ditto.pub"
]
],
"content": "All Chinese AI models are now switching to a combination of KV caching and sparse attention blocks. Huge speedups. This is https://www.minimax.io/\n\nDeepSeek 4 is also doing something like this. 1 billion tokens is now just 15$\n\nhttps://social.here.blue/media/7da6d2fc58670736c66f0c8bf571cbb7ee28c9399dd8df010d6376bb5a201181.blob",
"sig": "53c231e5b2fcd167592495200a7da6fe4c729705d27df91f6285c41d13aa2ef0ed46665b925e0b231bf65bd61887aa2de6965b752b898f778c621d75ac23cf2b"
}