​

Model Configuration

Model configuration describes which foundation model you fine-tune, how you adapt it, and how you deploy it. LLMTune supports 17+ model families across all major modalities.

​

Supported Model Families

LLMTune syncs with IO.net inventory and supports:

• LLaMA – LLaMA 3.3, LLaMA 3.4, and variants (Meta)
• Mistral – Mistral Nemo, Mistral 7B, and variants
• Qwen – Qwen3, Qwen-VL, Qwen2-Audio, Qwen2-VL-Video, and variants
• DeepSeek – DeepSeek R1, DeepSeek-Coder, and variants
• And more – 17+ model families total

Each model entry in LLMTune Models includes:

• Provider information
• Context length
• Parameter count
• Latency expectations
• Recommended use cases
• Deployment notes
• Evaluation metrics

​

Model Selection

​

Browse the Catalog

1. Navigate to LLMTune Models from the main navigation.
2. Browse the curated catalog of production-ready models.
3. Use filters to find models by:
   • Provider
   • Modality (text, vision, audio, code, etc.)
   • Size (parameter count)
   • Use case
4. Compare models side-by-side to find the best fit.

​

Model Comparison

• Side-by-side comparison – See multiple models at once
• Performance metrics – Compare latency, quality, and cost
• Usage guidance – Read recommendations for each model
• Deployment notes – Review deployment best practices

​

Adaptation Strategies

FineTune Studio supports multiple training methods:

​

Parameter-Efficient Methods

• LoRA / QLoRA – Fast iteration, lower compute costs
• SFT – Supervised fine-tuning with labeled data

​

Full Fine-Tuning

• Full fine-tune – Available for select models when deeper changes are required
• PPO / RLAIF – Reinforcement learning methods

​

Specialized Methods

• Code Generation – Optimized for code tasks
• Multimodal – Vision-language model training
• Audio methods – Audio understanding, ASR, TTS
• Embeddings – Text-to-embeddings training

​

Training Configuration

​

Hyperparameters

Key settings you can configure:

​

Compute Options

Choose your compute model:

• Traditional Computing – Single location, predictable performance
  • Single Instance or GPU Cluster
• Federated Computing – Distributed across global nodes
  • Privacy-preserving, unlimited scale, lower costs
  • Single Instance or GPU Cluster

​

Deployment Configuration

When deploying a fine-tuned model:

​

Version Control

• Version tagging – Tag each deployment with semantic versions
• Change logs – Document what changed in each version
• Approval workflows – Require approvals before promotion

​

Traffic Management

• Canary deployments – Gradually shift traffic to new versions
• Shadow deployments – Test without affecting production
• Blue/Green – Instant switch between versions
• Traffic splitting – Route percentage of traffic

​

Autoscaling

• Min/max replicas – Configure scaling bounds
• Timeout settings – Set request timeouts
• Resource allocation – Configure CPU/memory limits

​

Deployment Metadata

Deployed models store comprehensive metadata:

• Base model identifier – Original foundation model
• Dataset references – Which datasets were used
• Training configuration – Hyperparameters and method
• Training job ID – Link back to training run
• Endpoint URL – API endpoint for inference
• Current status – Active, paused, retired
• Version history – All previous versions

Use this metadata to:

• Reproduce runs – Recreate training configurations
• Audit model lineage – Track model evolution
• Debug issues – Understand what changed between versions
• Compliance – Document model provenance

​

Best Practices

1. Start small – Test with smaller models before scaling up
2. Use playground datasets – Validate your approach quickly
3. Monitor training – Watch metrics in real-time
4. Evaluate before deploying – Use LLMTune Evaluate to test models
5. Version everything – Tag and document all deployments
6. Plan rollbacks – Know which version to rollback to

​

Next Steps

• Browse the Model Catalog to see available models
• Learn about Fine-Tuning to train models
• Read the Deployment Guide to deploy models
• Check Model Configuration for advanced settings