Mint (MineLab)

Overview

Mint is the LoRA-centered model-training service attributed to MineLab in the supplied interview.

Current Profile

Mint emerged from MineLab’s attempt to train large models with limited H800 capacity. The source says the team found that LoRA could approach full fine-tuning on many tasks, then explored scaling base-model size, lowering adapter cost, and increasing the number of coexisting adapters.

Key Characteristics

• LoRA-based post-training service.
• Originated as a compute-constrained training approach.
• Targets low-cost specialization on shared foundation models.
• Explores serving many adapters with different capabilities and behavior.

Evidence

Training route

• 锐评国模与 Neo Lab，畅谈 RL 人生与 RLaaS | 对谈 Mind Lab 首席科学家骁腾 says the team adopted LoRA on a small H800 cluster and later organized the resulting work as Mint.

Scaling direction

• 锐评国模与 Neo Lab，畅谈 RL 人生与 RLaaS | 对谈 Mind Lab 首席科学家骁腾 names base-model scale, adapter cost, and adapter count as the three scaling axes.

Qualifications

• The source does not provide independent benchmarks, customer results, architecture details, or production serving scale.
• “Approaches full training” is task-dependent and remains a company-reported generalization.
• The generic product name is disambiguated as MineLab Mint in the canonical page key.

What Changed

• Established Mint as MineLab’s LoRA-based post-training and multi-adapter service.

Relationships

• MineLab - company developing the service.
• Ma Xiaoteng / 马肖腾 - source speaker describing the service.
• Low-Rank Adapter Model Diversity - model-society thesis enabled by many adapters.
• QLoRA - adjacent parameter-efficient training method.
• Model Post-Training Bottleneck - broader data, evaluation, and iteration constraint Mint still faces.