I run a self-funded solo AI research lab. I have 15+ years of experience developing and designing scalable solutions on the ServiceNow platform. By day I work on enterprise platforms. By night I run a local-first AI operating system on hardware I own, and I forge open models to run on it.

The cluster

It started as a pile of PCs under the desk and a mining-era stack of GPUs. Today it is sixteen machines and two NAS boxes, with a 40-gigabit RDMA fabric under the cluster core, a custom water loop, and a llama.cpp cluster carrying over four million tokens of live context across the serving fleet. It serves a 9b classifier lane, multiple 35b chat and agentic models, an 88b fast-chat model, and a 262b production heavy-tier MoE decoding at 29+ tokens per second, plus an internal vision summarizer, an embedding and reranking fleet, and an SDXL image generation box that does four images per pass.

The bench is busy too. DeepSeek V4 Flash (284B) went head to head with our own production 262b in a blind three-judge succession eval and swept it, 13 of 13 on the keyed probes where the incumbent took 9, and 2 of 2 on long-form prose where the incumbent scored none. Every one of the incumbent's losses was the same failure: exact-literal corruption at 84k context depth. It is now working through the promotion ladder, because winning an eval and being ready to carry production are not the same thing.

The models here are not drive-by uploads. They are built, quantized, and run daily on that cluster.

What we publish

The models

Newest first. Every model ships as a pair: the bf16 safetensors base you can re-quantize or fine-tune from, and the GGUF ladder you can actually run. If you are working on a single card, the 35B-A3B goes smallest: 11.2 GB at IQ2_XS, with 3B active per token.

Every published rung, plotted against size and bits per weight, lives on the catalog page. It is generated from the same build database that produced the files, so it cannot drift from what we actually shipped.

New: Qwen3.8-27B abliterated

Our first pass at a dense model. Everything else in this catalog is a mixture of experts that runs a fraction of its weights on any given token. This one runs all 27 billion every time, which makes it a different tradeoff rather than a smaller box: considerably more compute per token, and a different feel for it. It is also our first with vision.

Seven rungs, from Q8_0 at 29.05 GB down to IQ3_XS at 12.26 GB, with Q4_K_M at 16.84 GB as the volume rung. The f16 mmproj ships beside them: download it alongside whichever rung you pick, pass it with --mmproj, and you get the image path back. Without it you have a capable text model and no vision.

Two things we did not cut corners on. The vision tower survives abliteration intact. And the MTP block is pinned at or above q5_K on every single rung, never an I-quant, because no importance matrix covers the MTP head. An imatrix is collected during a forward pass that never runs the draft head, so there are no statistics for it, and a low-bit mixture applied blindly there produces garbage. If you re-quantize this model yourself, that is the caveat worth reading first.

The ethos

Every release ships with full provenance and credits the upstream base author. Own your models, your memory, your tools. Lean, efficient, no bloat, runs on anything. The same rule that runs the cluster.

Want a model treated? Open a discussion on our requests board. The stories behind these builds live at robinsonlabs.ai/writing.