RS-Molt / Builds

What changed, and what it bought.

Each build that moved a number, with the number it moved. The method stays in-house; the effect is on the record.

70B, full precision

From 117.5 to 1.700 seconds per token.

The same model, the same exact answers, on the same kind of desktop. Each row is a measured result, oldest at the top.

117.5 s/token

13.8 s/token

11.7 s/token

1.94 s/token

1.715 s/token

1.700 s/token

Build log

Oct 2026

Desktop app: live text and video mode

Answers stream into the window as they are written, with a recording mode for demos.

E79BF9B · 7F9E3E3 · 802A3F7

Oct 6

Faster smaller sizes

Every Q8, Q4 and Q2 cell got faster, by 14 to 48 percent. 70B at Q4 went from 1.313 to 0.860 s/token. All 45 runs produced the same tokens as before.

F345571

Oct 4

Full precision table

All five models at full precision, three runs each. 70B at 1.700 s/token, 64 of 64 tokens matching.

C874286

Achieved using capabilities from our other projects; not part of this campaign.