Measured
The result must come from an executable artifact and a recorded run.
DGL Tech / Evidence ledger No number
No number
without its conditions.
A benchmark without model, hardware, mode and date is marketing, not evidence. This ledger separates operational facts from lab results and future work.
Publication protocol 01 02 03 04
Scoped
Model, hardware, precision, workload and operating mode remain attached to the number.
Qualified
A lab result is never presented as a customer or production outcome.
Revisable
New measurements can supersede old ones, without rewriting history.
01 Lab
329 tok/s
Ternary decode on H100
Measurement context
Bonsai-8B Q2_0, H100 SXM5, deterministic single-stream greedy. Internal measurement from July 2026.
Mandatory qualifier
Lab benchmark; it does not represent a multi-user production workload.
02 Production
0
Python dependencies in serving
Measurement context
Inference is delivered by a Go binary with CPU, Metal and CUDA backends. PyTorch remains limited to training and export.
Mandatory qualifier
Support varies by architecture and modality.
The ledger is intentionally small. DGL Tech would rather publish three defensible facts than thirty numbers that cannot survive technical review.