Supported Models
MinT's base-model matrix has three categories: Qwen3 base models (available in both Community and Enterprise), Enterprise-only commercial models (GLM / Kimi / DeepSeek), and technically compatible same-generation open-source families.
Qwen3 base models (Community + Enterprise)
| Model ID | Recommended use |
|---|---|
Qwen/Qwen3-0.6B | Smoke tests, fast iteration (quickstart default) |
Qwen/Qwen3-4B-Instruct-2507 | Cookbook recipes (dapo-aime, chat-dpo, etc.), mid-scale chat |
Qwen/Qwen3-4B-Thinking-2507 | 4B-class reasoning / chain-of-thought tasks |
Qwen/Qwen3-30B-A3B-Instruct-2507 | Mid-to-large chat and instruction tuning (MoE) |
Qwen/Qwen3-235B-A22B-Instruct-2507 | Large-scale instruction tuning |
Enterprise-only commercial models
The compliant licenses for the following models are delivered with the Enterprise Edition and are not available in Community:
GLM-5GLM-5.1Kimi-K2Kimi-K2.5DeepSeek-V3
Technically compatible same-generation families
The following open-source families are architecturally the same generation as the supported models and can be onboarded on request:
- Qwen2.5, Qwen3 (Instruct / Thinking / Coder variants)
- Llama 3.x family
- Gemma 2.x and 3.x
- DeepSeek family
The full "coming soon" matrix on the website
The product website macaron.im/zh/mindlab/mint states support for "6 model families, 27 model variants," many of which are "coming soon." If you need to confirm concrete delivery dates before an enterprise purchase, contact sales.
How to choose
- No strong preference? Run a smoke run with any model in the lineup.
- Getting the flow working / fast iteration:
Qwen/Qwen3-0.6B. - Cookbook or mid-scale chat:
Qwen/Qwen3-4B-Instruct-2507. - Reasoning / chain-of-thought:
Qwen/Qwen3-4B-Thinking-2507. - Regular business (customer support / knowledge Q&A / document summarization):
Qwen/Qwen3-30B-A3B-Instruct-2507— noting that "a full fine-tune averages about 1 to 1.5 hours (30B model, A3B structure)." - Already licensed for GLM / Kimi / DeepSeek: pick the matching Enterprise-only model directly.
Switching the base model
MinT's consistent line is "switching the base model takes just one string change":
training_client = service.create_lora_training_client(
base_model="Qwen/Qwen3-30B-A3B-Instruct-2507",
rank=16,
)Script-level switching is also supported via the environment variable MINT_BASE_MODEL (for recipes that are MINT_BASE_MODEL-aware; see the docs site).
Note
The final list of available models depends on the license scope in your enterprise tenant contract. To onboard a new base model or a "coming soon" model outside your contract, contact sales at sales@mindlab.ltd.
What is MinT
MinT (Mind Lab Trainer) is Mind Lab's reinforcement-learning training infrastructure for large language models — a training platform accessed through a client SDK. You write a simple Python script on your own CPU machine; MinT manages the remote GPU cluster.
Quick Start
This chapter walks you through your first MinT training run in under 30 minutes: sign up, install the SDK, configure your API key, organize data, submit training, get a LoRA, and sample from it. Every step runs on remote GPUs — no local GPU required.