event

LLM Offloading Event

11 items · started 2026-10-03 · closed 2026-10-09

  1. moefit Run Qwen3.8-Flash-Next on a Mac that does not have enough RAM for the full model. Point your coding agent at this repo, give it the setup prompt below, and it can install and run a ~125B MoE model on Apple Silicon with 24–64 GB of u…

  2. The "Opus 5.5 makes videos" thing has been going around — and the interesting part is that it isn't generating pixels. The model writes a self-contained HTML/Canvas scene where every frame is a pure function of time, a headless browser cap…

  3. TL;DR: Native consumption of memory at the LLM level, no context. It's generally applicable to transformer-based models as well as Mamba and similar architectures.

  4. 24gb gddr7 2287 MHz 96gb ddr5 6400mhz core ultra 9 285hx, 5.5 ghz

  5. Hello guys! I've been out of the loop for a while.

  6. Hey all - I made a thing with my clanker Kevin: Ototo (https://ototo.dev/) - yes this is clanker driven dev. The idea is to offload code exploration and other token'y heavy tasks to a smaller model to do the work and chuck the result back.

  7. Jeff-Code A coding agent where a 0.8B decision model works alongside Qwen 3.8-27B. Coding tasks 47% faster (32% less time) on average, at the same pass rate.

  8. The surprise For a week I'd been tuning Qwen3.8-Flash-Next, a 125B-parameter MoE, on 🧠 yeti-cachy (my main homelab node: an i5-12600K with 64 GB DDR5 and an RTX 4070 12GB I paid $500 for). After a lot of flags and two experimental branches…

  9. could not extract summary

  10. Backburner Plug your iPhone into your MacBook with a 10 Gb/s USB-C cable and it helps run Qwen3.8-27B locally: Faster prefill (up to 64k context). For every batch of prompt tokens the Mac runs layers 1-40 and the iPhone runs 41-64 on its G…

← all threads