Open-weight defender: you don't need SOTA models 97% of the time
gnukeith · x · 2026-11-12
gnukeith pushes back on claims that open-weight models are doomed as US labs crack down on distillation: there are many other ways to gather training data. He argues benchmarks obsession is blinding people — Ernie 5.5, GLM 5.3 Flash and Qwen 3.8 27B each have their place, he personally finds DeepSeek V4.1 Flash and GLM 5.3 Flash more useful day-to-day than Opus or Astra, and 97% of tasks don't need SOTA models. He also credits open research from Chinese labs (DeepSeek R1, KV Cache efficiency, Xiaomi's MiMo RL training paper) as a boon to the whole ecosystem.
Related event: Open-source models remain viable as distillation fears dismissed(2 posts)→
More from Models
- One prompt, Blender MCP: Qwen4-Max builds a ramen stall at ~1/4 of DeepSeek's token cost — teortaxesTex · 2026-11-12
- 1.7B model shown driving browser and writing HTML autonomously, claimed to hit 4o-level benchmarks — JosephJacks_ · 2026-11-12
- repligate: I trust Mythos because it's aligned, not because it can't hack — repligate · 2026-11-12
- Baidu's ERNIE 5.5 Flash writes its own code to build editable CAD skyscraper model — heyshrutimishra · 2026-11-12
- Gemini 4 Carbon Spent 7.5 Hours in Robotics Lab and Broke a Gripper — ArtificialAnlys · 2026-11-12
- Humanity's Last Exam scores climbing sharply, Fable 5.2 tops chart at ~70 — SoulWriter23 · 2026-11-12