Adaptation log & releases
One post per model adaptation — what was added, what was verified, what to watch for — plus release notes and direction pieces. Adapting models is the day-to-day work of this project; this page is its public record.
v0.4.0 — images in, speech to text out
Vision models accept images through both inference APIs, and whisper.cpp joins the runtime behind POST /v1/audio/transcriptions — same lifecycle, same supervision. The catalog grows to 20 models.
Recipe schema v1 — models as data
The catalog stops being a Rust array. A recipe is one YAML file per model — repo, quant variants, VRAM requirements, engine args, a smoke test — and schema v1 just landed with validation and a published JSON Schema.
The base: any model, any GPU box, in ten minutes
Where OhMyGPU is going after v0.1 — and what the ten-minute promise actually costs, minute by minute, on a fresh GPU VPS. A budget, not a slogan.
Added to the catalog: Qwen3 4B Instruct 2507
Q4_K_M, 2.5 GB, native tool calling. Pulled, started and smoke-tested on Metal and CPU; comfortable on an 8 GB machine.
Gemma 3 joins the catalog: 1B, 4B and 12B
Three sizes from 0.8 to 7.3 GB, all Q4_K_M from ggml-org. No native tool-call parser in llama.cpp for this family yet — the flag says so honestly.
v0.1 — the runtime is the product
The first release: a headless daemon with two OpenAI-compatible inference APIs over one pipeline, a Management API for the whole model lifecycle, supervised llama.cpp, and a CLI that is deliberately thin.