AliceAI 80B-A3B at 262K: 6.0 GB KV. DeepSeek V4.1 Flash same window: 233 GB. 39x apart.
This is a dev post classified by Jev as Other (a tool drop), kept by the Dev Radar because it carries real work, not commentary.
AliceAI 80B-A3B at 262K: 6.0 GB KV. DeepSeek V4.1 Flash same window: 233 GB. 39x apart. Yandex dropped it today, Apache 2.0. 48 layers but 36 are KDA linear attention, no growing KV there. Only 12 layers do standard GQA, 2 KV heads at 256 dim. That is 1024 floats per token per layer, 12 layers, BF16: 24,576 B per token. At 262,144: 6.0 GB. Q4 weights about 45 GB. On a 128 GB Spark: 45 plus 6 is 51 GB at full context, 77 GB free. Active 3B at Q4 reads 1.69 GB per token. At 273 GB/s that is 162 tok/s roofline. My 27B at 18 is 9x slower on the same bandwidth. Haven't run it yet. MTP 1 layer i
Posted by 코지베어 🐻 CozyBear (1.2k followers) 1 days ago · 0 likes · 15 views · view the original post on X. Kept by the Dev Radar as Other.
More dev work like this
- we have chatgpt image 2.0 at home — @zainhas
- This greenscreen tool VFX artists have wanted for 30 years, solved at the pixel. — @0x0SojalSec
- 30 sec MiniMax H3 clip generated locally w/ a 4-step fused ref2va (omni) model in… — @blizaine
- Nano Banana 2.5 is incoming soon from @GoogleDeepMind. 🔥 — @mark_k
- You know whyyy this is crazy — @ItsmeAjayKV
- 🔥 ZastTranslate 1.21 is officially LIVE — powered by local Qwen-Image-2.1 in Pinokio… — @zast57
- it is mind blowing that you can now run an Adobe Photoshop-grade AI image model on an… — @cryptopunk7213
- google shipped 2 nano bananas this year and a 7b model just put both of them behind it. — @sudoingX
Every post is read and classified by Jev (TypeSafe): what it is, which market it belongs to, and whether the link is a real tool. 17k posts from 4.9k X accounts over the last 21 days, 1.9k tools, 12 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-20 22:16 UTC. Full method.