Qwengram-0.8B: I transferred Qwen3.8 Flash-Next’s n-gram memory into Qwen3.5-0.8B — 5.05% lower validation perplexity
For the last few days, I've been using two computers to run multiple agents. My 4x3090 machine is running Qwen 3.8 27B with parallel=2, so I can run two agents at the same time. My pi instances are running on a machine with 2x3060, running/managing smaller models such as Gemma 26B A4B or Qwen 35B A3B. Today, however, I needed my 4x3090 machine for some vLLM work, so I was missing my AI. I decided…
reddit.com ·
