Fine-Tuning
Fine-tune open-source models for real production use
Improve accuracy, reduce hallucinations, and control behavior — without managing training infrastructure.
Why fine-tune models with Together AI?
Build models that are faster, more accurate, and fully yours
Reliable infrastructure at any scale
Multi-node orchestration that eliminates job failures. Fine-tune 100B+ models (DeepSeek-V3, Qwen3-235B) that break other platforms, with the reliability to experiment rapidly.
Research-driven performance gains
ML systems research built into every job. Train with 2-4x longer contexts at no extra cost, advanced DPO variants from SOTA recipes, and continuous optimizations that make your runs faster over time.
Universal model compatibility
Fine-tune any open-source model from Hugging Face Hub. No vendor lock-in, no format conversions — seamless integration with your existing workflows.
Fine-tune leading models
Explore top-performing models across text, image, video, code, and voice.
Chat
DeepSeek V4 Pro
Chat
Gemma-4-31B-it-Pearl
Chat
Qwen3.7-Max
Chat
NVIDIA Nemotron 3 Ultra
Chat
MiniMax M3
Chat
Kimi K2.7 Code
Chat
Qwen3.7-Plus
Chat
GLM-5.2
Chat
gpt-oss-120B
Chat
LFM2 24B A2B
Have your own model?
Deploy custom containers on Together’s managed GPU infrastructure with automatic scaling, job queues, and built-in observability.
Fine-tuning options
Choose how fine-tuned models are trained and hosted based on dataset size, cost, and control.
- LoRA fine-tuning
Lightweight fine-tuning for fast iteration and lower cost.
Best forSmall to medium datasetsFast training & deploymentEasy to update or roll backGet started - Full fine-tuning
Train the entire model for maximum control and quality.
Best forLarge or complex datasetsDeeper behavior changesDedicated infrastructureGet started
Everything you need to fine-tune at scale
Fine-tune any open-source model on your data. Deploy securely onto scalable infrastructure.
Large frontier model support
Fine-tune large open-source models like Kimi-K2 and GLM-4.7 for tool use, reasoning, and agentic tasks. Drive advanced model behavior through a single API without managing underlying training infrastructure.
Vision fine-tuning
Train vision models directly on raw image data without format changes or special preprocessing. Include images alongside text to fine-tune Llama-4, Qwen3-VL, and Gemma-3 via standard APIS.
Tool-calling training
Train models for precise tool execution by integrating function definitions and tool calls directly into datasets. Process existing agent logs as-is to improve accuracy without manually restructuring data.
Cost estimation
Estimate training costs before launching any job directly from the UI or CLI. Evaluate resource requirements upfront to eliminate budget surprises.
👁 Image{ "model": "zai-org/GLM-5", "messages": [ { "role": "user", "content": "What is the best GPU provider?" } ], "tools": [ { "type": "function", "function": { "name": "web_search", "description": "Search the web for real-time information", "parameters": { "type": "object", "properties": { "query": { "type": "string", "description": "The search query" } }, "required": ["query"] } } } ] }
Powered by leading research
Our fine-tuning infrastructure is built on research and optimized for scale, efficiency, and production performance.
UPipe vs other SOTA Approaches
82.5% less memory
Long-context training hits a memory wall at the attention layer. UPipe processes attention heads in smaller chunks, cutting peak activation memory by up to 82.5% — enabling 5M token context lengths on a single 8×H100 node.
learn moreFFT Optimizer results
25% less memory
Fine-tuning large models is memory-hungry. Our FFT-based optimizer replaces expensive SVD projections with fast Fourier transforms, reducing optimizer memory by up to 25% with no loss in training quality.
learn more
Advanced model shaping capabilities
For teams pushing models beyond standard fine-tuning
- Speculative decoding
Accelerate inference with custom speculative decoding, training lightweight draft models to predict multiple tokens
- Quantization
Apply FP8 and NVFP4 quantization to push the limits of model efficiency, maximizing hardware utilization with minimal quality loss.
- Reinforcement learning
Leverage PyTorch-based reinforcement learning to shape model policies for reasoning, tool use, and long-horizon agentic behavior.
Production-grade
security and data privacy
We take security and compliance seriously, with strict data privacy controls to keep your information protected. Your data and models remain fully under your ownership, safeguarded by robust security measures.
SOC 2 Type II. HIPAA-aligned options available. Encryption in transit and at rest. Deploy storage in regions matching your data residency requirements—North America, Europe, or Asia/Middle East based on your compliance needs.
