Models - Monte Platform

A training config names its base model in model, as a Hugging Face ID. --model replaces it for one run:

monte train support-triage --config triage-sft --data support-replies \
  --steps 200 --from-base --model Qwen/Qwen3.5-9B

A new base model needs its own baselines before it trains. A run picks its GPUs by the model’s size, and --gpu and --gpus override the choice.

Supported Models

Every model below trains with SFT, DPO, and GRPO, and with distillation as a student when its family has a larger model with the same tokenizer. A size such as 30B-A3B is 30B total parameters, with 3B active for each token.

Choose a Model

Use case	Recommended models
A first run, or a fast iteration on the data	A model of 1B to 9B, such as `Qwen/Qwen3-8B`
Multi-turn tool use	An instruct model with a tool-call chat template: Qwen3, GLM, Kimi K2, or gpt-oss
Reasoning before the answer	Qwen3, DeepSeek R1, Kimi K2 Thinking, or gpt-oss
A distillation pair	A teacher and a student from one family, which share one tokenizer vocabulary: `Qwen/Qwen3-235B-A22B` and `Qwen/Qwen3-8B`