Models
Small models. One job each.
Trace
Draft models for one target.
A small model that predicts a large one, so speculative decoding accepts more of its guesses.
- v0
- 0.5B, distilled from Qwen2.5 7B. +3% accepted tokens.
- v1
- 35,000 prompts, trained on the 7B's own probabilities. In training.
- Bar
- +3 points top-1 agreement over the stock draft.
Lattice
Answers that survive 2 bits.
Models that keep their accuracy when compressed below 4 bits per weight.
- v0
- Calibrated Q2_K. GSM8K 21% → 42% at the same size.
- v1
- Quantization-aware training, and 7B to 32B on the Strix Halo.
- Bar
- At least 60% at Q2_K size, cheaper per answer than Q4_0.
Kinetic
A model shaped by energy.
Width, depth and vocabulary chosen from what each costs per token on real hardware.
- Status
- Planned, after Trace v1.
- Bar
- Beat a same-size Qwen on joules per correct answer.