Signal

3.93 GiB compression of Llama 3.1 8B. Under 6% repetition at 500 tokens where standard 3-4 bit quants hit 77-80%. Novel compression method, not ...

Coverage discusses speculative scenarios for 3-4; treat as market chatter and see linked sources.

reddit
beating_gepa
Evidence locked
Today's free sample is only available for the edition's flagship signal.
Evidence preview
  • 3.93 GiB compression of Llama 3.1 8B. Under 6% repetition at 500 tokens where standard 3-4 bit quants hit 77-80%. Nov...
    EVR-1 Maano
  • Squeezing a 14B model + speculative decoding + best-of-k candidate generation into 16GB VRAM- here's what it took (vi...
    Squeezing a 14B model + speculative decoding + best-of-k candidate generation into 16GB VRAM- here's what it took (via Reddit)
  • Beating GEPA/OpenEvolve/AlphaEvolve at a fraction of the cost (via Reddit)
    [R] LEVI