FINE TUNING

Train Your Own AI

Train, deploy, evaluate, and continuously improve custom models — all on one platform. No vendor sprawl. No infrastructure headaches.

Get started

Talk to our team

Train

Fine-tune with SFT, DPO, or RFT via Training Agent, Managed Training, or the Training API. Each loop starts here — with fresh data from the previous cycle.

Deploy

One-click deployment to an auto-scaling inference endpoint. Your model is live in seconds — no separate serving stack, no format conversion.

Monitor

Track latency, accuracy, and cost in production. Identify where your model struggles — the gaps become your next training dataset.

Collect

Gather new data from real production traffic: user feedback, failure cases, edge cases, tool-call logs. Feed it back into Train to close the loop.

TRAINING API

Customer success

Read Notion case study

"By partnering with Fireworks to fine-tune models, we reduced latency from about 2 seconds to 350 milliseconds, significantly improving performance and enabling us to launch AI features at scale. That improvement is a game changer for delivering reliable, enterprise-scale AI"

Sarah Sachs | AI Lead at Notion

"Fireworks has been a fantastic partner in building AI dev tools at Sourcegraph. Their fast, reliable model inference lets us focus on fine-tuning, AI-powered code search, and deep code context, making Cody the best AI coding assistant. They are responsive and ship at an amazing pace."

Beyang Liu | CTO at Sourcegraph

"The rLLM team is dedicated to pushing the boundaries of autonomous AI, which means our time is best spent on innovation rather than managing backend clusters. The Fireworks Training SDK lets us focus on our research instead of wrestling with infrastructure. The platform is fast, well-optimized, and just works."

Kyle Montgomery & Sijun Tan | Core Contributors, rLLM at rLLM

"By partnering with Fireworks to fine-tune models, we reduced latency from about 2 seconds to 350 milliseconds, significantly improving performance and enabling us to launch AI features at scale. That improvement is a game changer for delivering reliable, enterprise-scale AI"

Sarah Sachs | AI Lead at Notion

"Fireworks has been a fantastic partner in building AI dev tools at Sourcegraph. Their fast, reliable model inference lets us focus on fine-tuning, AI-powered code search, and deep code context, making Cody the best AI coding assistant. They are responsive and ship at an amazing pace."

Beyang Liu | CTO at Sourcegraph

"The rLLM team is dedicated to pushing the boundaries of autonomous AI, which means our time is best spent on innovation rather than managing backend clusters. The Fireworks Training SDK lets us focus on our research instead of wrestling with infrastructure. The platform is fast, well-optimized, and just works."

Kyle Montgomery & Sijun Tan | Core Contributors, rLLM at rLLM

"By partnering with Fireworks to fine-tune models, we reduced latency from about 2 seconds to 350 milliseconds, significantly improving performance and enabling us to launch AI features at scale. That improvement is a game changer for delivering reliable, enterprise-scale AI"

Sarah Sachs | AI Lead at Notion

THREE PRODUCTS, ONE PLATFORM

Start where you are. Go as far as you need.

Each product feeds into the same flywheel — train, deploy, improve, repeat.

Training Agent

Describe what you want. The agent does the rest.

For Product Managers & App Builders

From AI natives to enterprises, Fireworks powers everything from rapid prototyping to mission-critical workloads

Managed Training

Your data, your model, our infrastructure.

For ML Engineers

Use defined methods like SFT, DPO, and RFT. We provide reliable, scalable infrastructure for standard tuning with one-click deployment to production.

Training API

Own your model's intelligence.

For Advanced ML Platform Teams & Researchers

Full parameter tuning, 128k+ context window, custom loss functions, raw primitives. Push the frontier with complete algorithmic control on managed infrastructure.

	Training Agent	Managed Training	Training API
YOU BRING	Data + description of what you want	Formatted data + method choice	Your training loop + loss functions
WE HANDLE	Data prep, model selection, evals, training, deployment	GPU provisioning, distributed training, checkpointing, scaling	GPU execution, model parallelism, weight syncing, checkpointing, preemption recovery
ABSTRACTION	Fully automated	Method-controlled	Recipes → SDK → raw primitives
LORA / FULL PARAM	LoRA	LoRA and Full Parameter	LoRA and Full Parameter
PRICING	Per job, confirmed upfront	Per token / GPU-hr	Per GPU-hr

TRAINING METHODS

Three modalities, all feeding the same loop

Each method is purpose-built for a different stage of model maturity.

Supervised Fine-Tuning

Teach the model specific formats, tone, and domain knowledge from your labeled examples.

Classification & extraction
Tone & style alignment
Domain-specific Q&A

Reinforcement Fine-Tuning

Teach the model how to think, use tools, and solve multi-step problems with reward functions. The most powerful method for agentic AI.

Multi-turn agent workflows
Complex tool calling
Chain-of-thought reasoning

Preference Optimization

Align the model with human preferences and penalize undesirable behaviors without reward modeling.

Safety & guardrails
Hallucination reduction
Brand voice alignment

REINFORCEMENT FINE-TUNING

Improve model quality with reinforcement fine-tuning

Train frontier open-source models for multi-turn agents. User-friendly for app developers, powerful enough to surpass closed-source models in performance, quality, and tool interaction.

Technical deep dive

Train expert open models with just a few examples

RFT lets open models match frontier quality up to 10× faster, with just an evaluator and a few examples.

Superior multi-turn agent performance

Customize your model for complex multi-turn agents, dramatically improving performance with internal tools and function-calling.

No vendor lock-in, easy model swapping

Easily test with and train new models as the frontier evolves. Keep your same evaluator and pipeline, just swap the base model.

Enterprise-grade security and reliability

Choose flexible deployment methods, including full management by Fireworks or self-managed rollouts within your own environments.

SUPERVISED FINE-TUNING

Fine-tune with your own data

Customize model behavior by fine-tuning with your own data. Fireworks makes supervised fine-tuning fast, reliable, and cost-effective with an optimized training stack. Train large, state-of-the-art models using advanced methods like quantization-aware training to achieve ideal results.

Optimized training stack — Purpose-built for speed and reliability at scale.
Quantization-aware training — Train large models efficiently without sacrificing quality.
LoRA & full parameter — Choose the right tradeoff between speed and model fidelity.

Technical deep dive

PREFERENCE OPTIMIZATION

Align models to human preferences

Direct Preference Optimization (DPO) and ORPO let you shape model behavior using preference pairs — no reward model required. Ideal for safety, brand voice, and reducing hallucinations without complex RL setups.

No reward model needed — Simpler pipeline than RLHF — just chosen/rejected pairs.
Safety & guardrails — Penalize harmful or off-brand outputs directly.
Hallucination reduction — Teach the model what not to say, not just what to say.

Technical deep dive

WHY FIREWORKS

The only platform that combines full-spectrum training and built-in inference

Serve personalized models at scale

100s

Models per GPU

1-click

Deployment

Zero

Extra infra cost

ALTERNATIVE	EXAMPLES	THE LIMITATION	FIREWORKS ADVANTAGE
Closed Models	OpenAI, Anthropic	No weight ownership. High cost. Zero portability. No retraining loop.	[check] Open-source models you fully own. Retrain and redeploy continuously.
Training-Only	Fragmented vendors	Train here, serve elsewhere. Every iteration pays a migration tax.	[check] Unified platform. Training completes → model is live → collect data → retrain.
Cloud-Native	AWS, GCP	Training and inference are separate silos. No open model expertise.	[check] Model-agnostic. 1-click hot-loading from training to inference.

START BUILDING TODAY

Start the flywheel today

From first experiment to continuously improving production models — all on one platform.

Get started

Talk to an expert