Models

What I publish

Requants, hardware-specific builds, and my own fine-tunes of the models I run myself, published on HuggingFace. Mainline-compatible tiers and my single-GPU Gyro builds so far; more will land here as I build and quantize them.

Illustration of a llama peeking out of a small server box, with a glowing halo above its head

Gyro 3.8 Flash Next

Qwen3.8-Flash-Next (125B MoE) on one GPU. Gyro-S fits a 32 GB card at 27.6 GiB; Gyro-M matches a 43.8 GiB 3-bit build with 20% less memory. Built with our own rotor code.

Signal 3.8·27B

Qwen3.8-27B that gets to the answer faster. 57% fewer answer tokens and 52% fewer thinking tokens than the base model, at matching or better answer quality.

Qwen3.8-27B AP

High-precision GGUF requants of Qwen3.8-27B, measured by KL divergence against the unquantized model. Beats size-matched ISTA and Unsloth builds on every tier I've tested, pushing the size-vs-fidelity pareto frontier out further than either.

Qwen3.8-Flash-Next

Seven mainline-compatible tiers plus a Strix Halo build, each benchmarked against the leading community builds at matching sizes. My smallest tier beats the field in its class.