Service Pricing¶
Plan training and inference costs by model and usage, and choose the resources that fit your workload.
Mint Recursive provides model training and inference services, including supervised fine-tuning (SFT), reinforcement learning (RL), and deployment of trained models. The table below lists usage-based rates. Contact sales for dedicated GPU resources and custom services.
Billing overview¶
Prices are in USD per million tokens.
| Billing item | Description |
|---|---|
| Prefill | Processing input tokens during prefill. |
| Cached Prefill | Processing input tokens that hit the cache. |
| Decode | Generating output tokens. |
| Train | Processing tokens during model training. |
Training and inference are priced separately. Select the rate for your model, training method, and billing item. For ordinary hosted-model API inference prices, refer to Models & Pricing in the console.
Model list prices¶
| Model / training method | Architecture | Parameters | Prefill | Cached Prefill | Decode | Train |
|---|---|---|---|---|---|---|
| Qwen3.8 SFT | Dense | 27B | $1.860 | $0.372 | $5.595 | $4.103 |
| GLM5.2 SFT | MoE | 744B | $5.000 | $1.000 | $14.000 | $18.500 |
| GLM5.2 RL | MoE | 744B | $5.000 | $1.000 | $14.000 | $24.050 |
| GLM5.3 SFT | MoE | 744B | $5.000 | $1.000 | $14.000 | $18.500 |
| GLM5.3 RL | MoE | 744B | $5.000 | $1.000 | $14.000 | $24.050 |
| DeepSeek V4.1 Flash | MoE | 552B | $0.600 | $0.120 | $3.600 | $4.980 |
Dense denotes a dense model, MoE denotes a mixture-of-experts model, and B denotes billion parameters. SFT stands for supervised fine-tuning; RL stands for reinforcement learning. Training methods in the model labels distinguish the applicable rates.
Dedicated GPUs¶
Dedicated GPU pricing for every model listed above is TALK TO US.
For dedicated inference resources or a custom deployment, contact sales with your target model, training or inference use case, expected usage, concurrency, and deployment requirements to discuss resource configuration and pricing.
Custom services and quotes¶
For LLM as a judge, large-scale training, and other custom services, contact sales@mindlab.ltd. For judge workloads, include your evaluation rubric, data volume, and intended use of the results. Service scope and fees are subject to the agreed quote or order.
Ready to begin? Read the Quickstart or explore the Training and Deployment guides.