TQ4
Stores weights in 4 bits with a learned scale and codebook while retaining trainable controls above the compact substrate.
- Measured
- 870 MB and 0.9998 cosine.
- Missing
- Full method publication.
TQ4 on Qwen3-1.7B
One internal run cut linear-weight memory from 2.7 GB to 870 MB while the top answer matched on five prompts.
- Output
- 0.9998 cosine
- Runtime
- 238 tok/s
Qwen3-1.7B output comparison
0.85 out of 1.00
bf16 comparisoncompact model
Measured cosine scores: 0.85, 0.91, and 0.9998.
Method status
Each method below separates what ran from what remains unpublished.
Native FP4
Uses hardware 4-bit math without expanding every operation back to 16 bits.
- Measured
- Qwen3-8B at 40 tok/s and Qwen3-14B at 21 tok/s on 16 GB.
Megakernel
Places inference and training in one small program.
- Artifact
- About 2,500 lines of Rust and 1,300 of CUDA.
- Missing
- Full method and evaluation.
QGRE
Rewards a correct model attempt instead of training only by copying a completed answer.
- Artifact
- Internal Hamiltonian derivation on one 16 GB GPU.
- Missing
- Full method and token-level evaluation.