README.md

---
license: apache-2.0
tags:
- requests
- gguf
- abliterated
- reap
- quantization
---

# Request a quant or abliteration

Robinson Labs forges open models on a real homelab cluster: abliterated builds, REAP expert-pruned
MoE variants, importance-matrix (imatrix) GGUF quant ladders, and custom cluster-fit quants. This
repo is where you ask for one.

See what we have published: [huggingface.co/RobinsonLabs](https://huggingface.co/RobinsonLabs).

## How to request
Open a new **Discussion** (the **Community** tab at the top of this repo) with:
- A link to the model on Hugging Face (the original or base repo).
- What you want: a GGUF quant ladder, an abliterated build, a REAP-pruned variant, or a fit-quant
  tuned to a specific VRAM and context budget.
- Your target, if you have one: VRAM size, context length, or a specific quant (for example IQ4_XS,
  Q6_K).

One model per Discussion keeps things easy to track.

## What we prioritize
We are a one-person homelab, not a quant farm, so we bias toward the gaps, the builds that do not
already exist:
- **Abliterated builds** where there is no good uncensored variant yet.
- **Abliterated REAP / expert-pruned MoE**, a niche almost nobody fills.
- **Custom cluster-fit quants** tuned to a real VRAM and context budget (attention-path precision
  kept high where it counts, experts run lighter).
- Models that lack a solid imatrix GGUF ladder.

If a model already has good quants from the usual sources, we will likely point you there. Where we
add value is the build that is missing.

## What to expect
Best-effort, on hardware we own. No SLA, no queue promises. Every release ships with full provenance
and credits the upstream base author. On abliterated builds, harm guardrails stay intact by design
(self-harm prompts still redirect to help, and the build is not meant to assist genuine wrongdoing).

You are responsible for your use of any model published here.