One 36B MoE. Six EXL3 releases. GPUs from 8 GB to 96 GB. 🔥
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.
One 36B MoE. Six EXL3 releases. GPUs from 8 GB to 96 GB. 🔥 K2-Horizon-MoVA-36B-A4B is now local for almost everyone!! 👇🏼 2.5 bpw: 13.27 GB with 83.7% top-1 agreement 8.0 bpw: 38.07 GB with 96.43% agreement, at roughly half the BF16 weight footprint This is exactly why I use EXL3 when the goal is maximum model quality per GB. 𝗞𝟮 𝗛𝗢𝗥𝗜𝗭𝗢𝗡 𝗜𝗦 𝗔 𝗩𝗘𝗥𝗬 𝗜𝗡𝗧𝗘𝗥𝗘𝗦𝗧𝗜𝗡𝗚 𝗠𝗢𝗗𝗘𝗟 It stores 36B parameters but activates only about 4B per token through a combination of: → Mixture-of-Experts feed-forward routing → Mixture-of-Values attention → 512K native context → fully open weights, training data, l
Posted by Cruz (2.1k followers) 1 h ago · 10 likes · 623 views · view the original post on X. Kept by the Dev Radar as Other. Tools mentioned: k2-horizon-mova-36b-a4b-exl3, k2-horizon-mova-36b-a4b-exl3, k2-horizon-mova-36b-a4b.
More dev work like this
- Opus 5.5 is an excellent model! I benchmarked it on induction, and it comes a clear… — @s_batzoglou
- Agentic AI is live on CookMyMeme. 🤖🍳 — @CookMyMemeCoin
- 想找一个专门针对 A 股的 AI Agent 项目，可以看看这个。 — @bkdgiffug
- Good morning to all Commemorates reaching 10 million users and displays core ecosystem… — @LokeshP95013
- NEW UNCENSORED Xiaomi MiMo-V2.6-Flash model you can run locally with Zero refusals. — @0x0SojalSec
- 🔥 LightMem-Ego just got a major upgrade — new system, faster interaction, and a new… — @zxlzr
- Ok ….progress update — @u1tra_instinct
- Locally Laya just beat cloud Jev at live Tetris. — @0x0SojalSec
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. 21.9k posts from 5k X accounts over the last 21 days, 2.5k tools, 12 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-23 04:29 UTC. Full method.