Part 2 / correction on TwIL-8B from @thewebAI.
This is a AI post classified by AGK Intelligence as Open & local models (a model release), kept by the AI Radar because it carries real work, not commentary.
Part 2 / correction on TwIL-8B from @thewebAI. Yesterday’s run looked slow and incomplete. That was on me. TwIL’s KV cache needs are higher than most models in this size class. Same 51.2K window, but more KV bytes per token than peers like Ornith/NeoHorse, so my usual load spilled the cache into system RAM instead of keeping it on the 10GB card. S1 sat around 16 tok/s with thinking off and about 6.6 with it on. PA with thinking on never finished cleanly. I fixed the load so KV stays on the GPU (51.2K context, 41/41 layers, Flash Attention, q4_0 K/V). Re-ran the suite today. Corrected speed
Posted by GooGZ AI (1.6k followers) 2 h ago · 9 likes · 661 views · view the original post on X. Kept by the AI Radar as Open & local models.
More AI work like this
- Qwopus3.8-27B-Coder-Flash can fit in an RTX 3090 — @melvindvivas
- BAIDU JUST DROPPED A MAJOR RELEASE FOR OPEN-WEIGHT AI — @DataChaz
- Meet Meerkat2-14B-GGUF: a 14B medical AI that reasons in Korean and English. Built for… — @HuggingModels
- Inkling 2 Small is currently the top trending work on Papers with Code, congrats Mira… — @NielsRogge
- Most people think running Local AI starts with downloading a 27B parameter model. — @javierbits
- Weird easter egg in the Qwen launch image — @bijanbowen
- Meet Qwen3.8-27B-Thinking-SecOPD: a 27B image-text-to-text model built to resist prompt… — @HuggingModels
- By the way, I love these charts on Slop theme on Priors — @AdamHoltererer
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 65.7k posts from 5k X accounts over the last 14 days, 10.4k tools, 19 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-11-06 13:30 UTC. Full method.