DoctorShotgun is a public pseudonymous developer and language-model fine-tuner. Their work includes the Magnum-Diamond models for creative writing and practical guidance for running large mixture-of-experts models across CPU and GPU memory.[1][2][3]
Start with Local MoE inference with CPU and GPU memory →
Links
Hugging Face
GitHub Contributions
Fit sparse models across different memory types
A guide co-authored with Geechan explains why a mixture-of-experts model’s total parameter count differs from the number used for each token. It places always-active components on the GPU while keeping sparse routed experts in CPU memory, then tunes the split and batching for the available machine. The useful result is a measurement process, since a command optimized for one hardware configuration may perform poorly on another.[2]
Fine-tune models for creative writing
The 24B Magnum-Diamond model adapts a text-only Mistral Small 3.2 checkpoint with an rsLoRA adapter. Its card describes changes to tokenization and loss masking, along with prompt templates and sampling guidance. The stated objective is creative prose; those preferences and the model’s name are not an independent evaluation of writing quality.[3]
Support collaborative releases
The Anthracite Magnum v4 card separately credits DoctorShotgun for sponsoring the compute used in full fine-tuning. Compute sponsorship supports the release without implying sole authorship of the model.[4]
Tools, writing & talks
Writing, research & talks
Start here
Local MoE inference with CPU and GPU memoryhuggingface.coA practical guide to allocating model components and measuring local inference performance.
Begin with the MoE component diagram and memory budget, then tune using measurements on your machine. You need a compatible llama.cpp build, an appropriate GGUF model, and enough combined RAM and VRAM including cache overhead.
Co-author with Geechan[2]
Tools
- MS3.2-24B-Magnum-Diamondhuggingface.co
A creative-writing adaptation of Mistral Small with published prompt and adapter details.
Model fine-tuner[3]