Xyntetik Shade research, experimental
Shade discovers. Runner productizes.
Shade is the private research lab of the Xyntetik program: experimental work on how models can be made smaller, faster or more inspectable without lying about what that costs. It is not a product and nothing in it is finished. Its outputs are findings, and where a finding survives, a mechanism the Runner engine can ship with a gate.
How it works
Two missions, one law.
A methodology lab that proves training, quantization, pruning and evaluation methods on governed data, where the bar for any work is a publishable method or a consulting proof-point. And a discovery queue of architecture and mechanism research that the engine can exploit. Training a competing model from scratch is not a goal; it appears only as a control, at toy scale, in service of a finding.
The law is a written charter adopted 2026-08-27: preregistration before measurement, a split between discovery and publication reproducibility, corrections recorded in place, and operational rails winning over epistemic enthusiasm. Every study owns its data, configs, scripts and logs, and a report whose closed sections are immutable.
What the bytes actually buy: the memory-versus-capability boundary across local models, and whether a byte partition predicts the kind of capability.
Removing parts of a model and measuring what it costs against the parent: expert pruning, decoder-depth removal, attention-sublayer removal, and healing after the cut.
Mechanisms the engine could exploit: draft sources, precision placement, context surgery. Measured first in the lab, shipped in Runner only behind a gate.
Evaluation methodology, including a running record of where the house fidelity bar does and does not transfer to public benchmarks.
Public outputs
What has left the lab, with its numbers.
Shade's repository is private. What it publishes are artifacts and measurement reports on Hugging Face, each with provenance and the measured envelope, and engine features that graduated from a finding.
- passes barThe decoder cut and healed, then quantized by Runner's own quantizer: 4.75% removed in the Q4_K file (14.61 GB), 6.34% in the Q8_0 (27.58 GB). Both pass the fidelity bar against the unmodified BF16 parent on a held-out split: margin-qualified top-1 97.74% and mean KLD 0.04810 for the Q4_K, 98.34% and 0.03883 for the Q8_0. The Q4_K cost was predicted to within 0.6% before the file was built, from the measured independence of quantization noise and surgery error. The report also records where the fidelity bar does not transfer to public benchmarks (measured 2026-08-29), rather than leaving that to be discovered.
- published 2026-09-04A study of Google's 31B model found one attention sublayer that can be dropped at KLD 0.0223 on 44,413 held-out positions. The cut frees 64 MiB of KV cache at 4k context and 512 MiB at 32k. It became the first sublayer-removal artifact, and the mechanism became an engine feature (below).
- shippedGraduated into Runner:
--remove-sublayerRunner v0.4.7 (2026-09-04) physically drops a block's attention or dense FFN tensors and declares the absence in the GGUF header using llama.cpp's own per-block convention. The engine omits the branch and reserves no KV rows for it; the release carries a bit-identity gate against the study's zeroed form. CPU path, dense blocks only; llama.cpp refuses the file by name, and the card says so. - negativeFrontier reports where nothing beat upstreamThe Lightning-30B prune frontier (keep-126 passes at 99.50% and KLD 0.026; the 1.37% saving was not worth an artifact) and the Muse-Glimmer quant frontier (Meta's own Q4_K_M passes the bar; six plans measured, none beat it). Published as reports, no weights, stated openly.
An artifact that passes the bar is a measured derivative with its envelope attached, not an endorsement that it is better than the original for your workload. A report is a measurement, not a product. Work inside Shade that has not been published does not exist as far as this site is concerned.