Products

FORGE

Mixed-precision quantization, chosen by the Hessian

Sub-3-bit quantization usually stops a model being a model. FORGE is a post-training pipeline that decides per tensor rather than uniformly: it measures Hessian flatness to find the weights that cannot survive ternary, holds those at higher precision, and crushes the rest. On Falcon-H1-7B that is a 4.4x reduction to 3.48 GB with factual recall intact.

Precision map across 44 blocks showing most tensors at TQ2_0 with ffn_down and ssm_out held at Q6_K, above a Hessian flatness plot where ssm_out spikes to 6.78 against roughly 0.33 elsewhere.
2.06
bits per weight, average
4.4x
smaller than F16
12/12
factual probes recovered

The failure that motivated it

Our first pass quantized every reachable tensor to ternary. The result was not a loss of fluency, which would have been easy to spot. It stayed articulate and became confidently wrong.

probeuniform ternarymixed precision
carrot cake ingredients"carrot cake mix, carrot cake mix, ...""Carrots, sugar, flour, eggs, butter, cinnamon..."
capital of Australia"Sydney""Canberra"
boiling point of water"99 C""100 C"
Only the first is a visible loop. The other two read as authoritative.

What the Hessian showed

Across the reachable tensors, outlier flatness clustered around 0.33 — except for ssm_out, the projection that writes the Mamba-2 scan's output, which measured 6.78. A 20x outlier, and structurally explicable: it sits immediately after the scan, in the same position ffn_down occupies after the SwiGLU, where a fusable rotation cannot reach it. Crushing that tensor was destroying the path the model's factual recall travels through.

Excluding ssm_out and ffn_down from ternary and holding them at Q6_K costs 0.23 GB. Mean flatness across the still-quantized set drops from 1.1881 to 0.3818 — the outlier is simply no longer being crushed.

Measured

buildsizepplvs F16factual probes
F1615.18 GB6.58751.00x12/12
uniform ternary3.25 GB10.75901.63x9/12
mixed precision (shipping)3.48 GB9.16331.39x12/12
Wikitext2 perplexity against the F16 source.

Honest limits

The shipping run changed two variables at once — preserving ssm_out and raising calibration from 32 to 128 sequences — so the recovery cannot be attributed between them. The flatness evidence makes ssm_out the likelier cause, but that is inference, not measurement. A third run would separate them.

  • No same-size baseline. A standard Q4_K_M build has not been measured, so "better than conventional 4-bit at this footprint" is not a claim this data supports.
  • Cross-turn repetition persists at roughly 17% over four turns even with a correctly sized penalty window. It is reduced, not eliminated.
  • Meta-instructions fail. Direct instructions are followed; instructions about the conversation are not. That is a quantization ceiling, not a templating bug.
  • Evaluated with targeted factual probes and perplexity, not MMLU or HellaSwag.