Fine-tune
Task priorities
Chat quality ≈ human preference; Documents and Instruction following are derived from benchmark baskets.
Training hardware
1 × 24 GB = 24 GB usable
Inference hardware
Drops models whose Q4 quant won’t run on the serving hardware.
Recipe
DPO adds a reference model; RL (GRPO-style) adds rollout memory and compute.
Training method
Dataset size
Training budget
License
Total parameters
Context window
Architecture
Organization
Required modalities
Open-weight models only — closed API models can’t have their weights tuned; hosted fine-tuning APIs are out of scope.
Which model should you fine-tune?
Open-weight models ranked for your constraints — training VRAM per method, estimated cost, and task quality. Formula-derived estimates: see methodology.
211 fit your hardware · of 345 trainable
ShowSort
#ModelWeightsLicenseParamsMethodTrain VRAMCostQuality