Qwen
seriesFull-spectrum open-source model family with dual-mode reasoning across 119 languages.
Mind Lab Toolkit
MinT
MinT is Mind Lab's enterprise RL infrastructure for LoRA post-training, turning real product experience into reproducible training and sampling loops so models keep improving in real business scenarios.
Local verification -> cluster execution -> online learning loop
DATA
Enterprise data, feedback, test sets
Model
Full range of open-source models
ALGO
SFT, DPO, RL and more
Training
Cluster orchestration, checkpoints, eval
Sampling
Online feedback, replay, iteration, local deployment
Post-training as an enterprise loop.
MinT abstracts compute scheduling, distributed rollout, training orchestration, sampling updates, and evaluation replay into a reproducible system for learning from real product feedback.
- Local verify. Teams validate data, targets, and loss design before cluster execution.
- Cluster train. MinT orchestrates LoRA post-training, evaluation, checkpoints, and replay.
- Online learn. Product feedback becomes the next iteration of domain model behavior.
Supported models
Supported model families.
Community edition includes Qwen by default; enterprise edition provides access to additional open-source model families.
GLM
seriesBest-in-class Chinese language capability with strong tool-calling and code generation performance.
MiniMax
seriesHigh-efficiency MoE model family excelling in coding and agentic workflows with native million-token context.
Kimi
seriesTrillion-parameter MoE architecture built for long-horizon autonomous coding and multi-agent orchestration.
DeepSeek
seriesPioneered pure-RL reasoning emergence with standout performance on mathematical and code reasoning tasks.
Plans
Customized solutions for enterprise teams.
Community edition works out of the box; enterprise edition adds higher availability, broader model coverage, and private deployment.
Capability
Community
Enterprise
Support
Community and self-serve onboarding
5*8 enterprise support response
Included usage
5M token quota, covering 0.6B–235B models
Custom quotas and reserved capacity
Model access
Qwen series
Qwen, GLM, MiniMax, Kimi, and DeepSeek series
Workspace
Standard cloud workspace
Private deployment, enterprise VPC, hybrid deployment
Training loop
Basic training loop and docs
Dedicated solution design and rollout support
Security
Standard safeguards
Security review, compliance support, audit logs
Security & compliance
Enterprise controls for sensitive training data.
MinT is designed for teams that need model improvement without giving up privacy, data ownership, deployment control, or auditability.
Privacy & security
Secure by default.
Enterprise data is not used to train shared foundation models by default. MinT supports sensitive-data detection, redaction, minimization, encryption in transit and at rest, least-privilege access, audit logs, and continuous monitoring.
- PII and sensitive-field handling before training
- Full-link encrypted transfer and encrypted storage
- Role-based access, logs, and review-ready controls
Data ownership
Enterprise data belongs to you.
All Enterprise Data you upload to MinT (including training data, training outputs, and inference inputs and outputs) remains wholly owned and controlled by you. MinT will not use Enterprise Data to train its own models or for any other commercial purpose without your prior express authorization. Cloud workspaces are tenant-isolated, with support for dedicated cloud, enterprise VPC, private, and hybrid deployments.
- Enterprise-owned datasets, LoRA weights, and reports
- Cloud multi-tenant isolation and private workspace options
- Dedicated cloud, VPC, private, localized, or hybrid deployment
Compliance
Built for regulated teams.
MinT complies with China MLPS Level 3 requirements and supports enterprise obligations under the Cybersecurity Law, Data Security Law, Personal Information Protection Law, and industry audit programs. Compatible with multiple chip platforms for localized training deployment.
- China MLPS Level 3 compliance support
- Audit support for domestic data and privacy obligations
- Multiple-chip compatibility for localized training deployment
Cases
Helping models evolve with the business.
The following cases are distilled from MinT solution materials and enterprise product overview.
Healthcare
Medical Coding
Medical coding and case-record post-training for quality-control workflows across clinical scenarios.
-90%GPU cost
27 case-record QC indicators
8,000+ physicians referenced in rollout
Training steps reduced by 50%
On-policy training time reduced by 50%
Personal Agent
Personalized Agent
Personal agent infrastructure where many 1T model instances share one base while each user keeps LoRA differences.
340Kmini-apps
1M users in the case deck
Only 5B LoRA diff-params per user
Thousands of 1T instances supported
Deep memory-driven personalization
FinTech
Smart Customer Service
LoRA post-training for a leading fintech support workflow with compliance, takeover, and cost improvements.
+34%CSAT
18M-user service scenario
Compliance accuracy 89% to 99.2%
Human takeover 38% to 15%
Training cost reduced by 68%
Government Cloud
Citizen Services
Policy-aware service agents for public cloud scenarios with fast policy refresh and measurable transfer reduction.
6hpolicy updates
23M citizens served
1,400+ services covered
Policy error rate 23% to 5%
Agent transfer 44% to 21%
Start the loop
Turn enterprise experience into model parameters.
Register, prepare your enterprise data, define the target, connect MinT, and receive reproducible training scripts, evaluation reports, deployable LoRA weights, and the next iteration plan.
Start the loop