Sixteen machines and two NAS boxes, a 40-gigabit RDMA fabric under the cluster core, and a llama.cpp cluster carrying over four million tokens of live context. Everything published here is built, quantized, and served on that cluster. Not drive-by uploads.
Bits per weight against file size, for all 72 published quants. Straight out of our provenance database, measured from the built artifacts rather than estimated from the parameter count on the label.
Every model ships as a pair. The bf16 safetensors base is the master you re-quantize or fine-tune from. The GGUF ladder is what you actually run.
The parts that are ours, and the parts that are just done carefully.
Single-direction weight orthogonalization that relaxes the hard-refusal reflex while keeping harm guardrails intact by design. We do the surgery in-house from the base model rather than re-cutting somebody else's abliteration.
Importance matrices calibrated on our own corpus, not a generic wikitext pass. The card for each ladder names the corpus it was calibrated on, because a quant is only as good as the activations that guided it.
We abliterate community REAP expert-pruned MoE variants and publish the pair. Pruning and abliteration interact, and doing both in the wrong order costs you the model, so the lineage of every rung is recorded.
The multi-token-prediction head survives quantization. It also gets pinned: no importance matrix covers that block, because the calibration pass never runs the draft head. Left alone it gets quantized blind. We hold it at or above q5_K on every rung and say so.
On multimodal bases the vision tower survives abliteration, and the f16 mmproj ships beside the ladder. Download it alongside whichever rung you pick and the image path works. Without it you have a capable text model and no vision.
Recipes cut to a specific VRAM and context budget instead of a generic rung. Attention path precision stays high where it counts, the experts run lighter. These come out of actually needing to fit a model onto our own boxes.
The engine for this already exists. WebGPU, Transformers.js, weights cached in your browser, everything running on your own hardware with nothing sent to a server. We built and benched it.
What is missing is a model of ours small enough to run in it. The limit is not really the GPU, it is the download: a visitor will wait for something around a gigabyte, which in practice means a model of a few billion parameters. Our smallest is 27 billion. We could drop somebody else's small model in here and call it a demo, but a showcase of another lab's weights on our own page is not worth having.
So a small one is being built. When it lands, this panel turns on, and what you talk to will be ours end to end.