I build and tune local AI: inference engines, model recipes and dev tools, so large models run fast on hardware you own. Current focus: NVIDIA DGX Spark (GB10).
| Recipe | What it is |
|---|---|
| Nemotron-3.5-Super-VL abliterated | NVIDIA's 120B vision-language model, uncensored, on TensorFold or vLLM; refusals from 98 % to 2 %, tasks and speed unchanged |
| Qwen3.8-Flash-Next | 262,144-token windows, three streams, NVIDIA's public FP4-Local checkpoint |
| DeepSeek-V4-Flash | Engine patches and a Docker image for the 0731 model on a single Spark |
| Model | What it is |
|---|---|
| Nemotron-3.5-Super-VL-120B-A12B-NVFP4-abliterated | The uncensored build behind the recipe above, in NVIDIA's own tensor formats |
| Qwen3.8-27B-NVFP4-FP8-Mixed-LH | A mixed NVFP4 / FP8 quantization for single-GPU Blackwell machines, SGLang and DFlash ready (uncensored variant) |
Contributions to TensorFold
All of them, open ones included: my TensorFold pull requests.
More projects and tools will appear here as I open-source them.
🤗 Hugging Face · X · Ko-fi