Open Source

Community pleads for Qwen3.8 in smaller sizes, not trillion-parameter giants

Most users can't run 2T+ models, but 27B to 397B would work on today's hardware.

Deep Dive

A Reddit user argues that instead of releasing 2T+ models, continuing to develop highly capable small to medium LLMs would keep the community vibrant, since almost no one can run the massive 1.5-2T+ beasts. The user points out that smaller models, especially with CPU expert offloading, can run comfortably on today's wide range of systems. They criticize Chinese labs for trying to match frontier-class trillion-parameter open-weight models, saying this trend doesn't help the local model community innovate—it just gives big corporations a cheaper alternative to commercial frontier models.

Key Points
  • Requests Qwen3.8 in 27B, 35B, 122B, and 397B parameter variants for broader accessibility.
  • Argues trillion-parameter models exclude most users and only serve large corporations.
  • Highlights that CPU expert offloading makes smaller models viable on current consumer hardware.

Why It Matters

Smaller open-source LLMs keep AI development accessible, preventing centralization of power in big tech.

📬 Get the top 10 AI stories daily