Open Source

Reddit asks: Can a 397B model run locally on 256GB RAM?

Qwen 3.6 skipped local release; users seek 397B competitors that fit in 256GB...

Deep Dive

A Reddit user asks if any model can run locally to compete with Qwen 3.6's 397B-17B variant, noting that this version of Qwen was not released for local deployment.

Key Points
  • No open-source 397B MoE model currently confirmed to run on 256GB RAM without extreme quantization.
  • Llama 3.1 405B at 4-bit uses ~200GB, making it the closest option but still ~95GB over budget.
  • Qwen 3.6 (community variant) omitted its 397B-17B local release, leaving a gap in ultra-large local models.

Why It Matters

Highlights the tension between model size and local deployment; drives demand for efficient MoE architectures.