# SynqTune > SynqTune trains small open language models on a company's own examples, inside the company's own AWS account, and compares the result with the LLM they use today (such as GPT or Claude) on the same held-out examples before they switch. Data and model weights stay in the customer's S3; training runs on EC2 instances in their account. Key facts: - Runs in the customer's AWS account through one IAM role with an External ID; deleting the role revokes access. - The role is scoped: only its own S3 bucket, only EC2 instances it tags `synqtune:managed=true`, provider keys it can save but not read (only the customer's instances can), and stack updates it can propose as change sets but not execute. - A CloudWatch alarm in the customer's account terminates any SynqTune instance idle for an hour, even if SynqTune is down. - Fine-tunes open models (typically 1B–8B parameters, e.g. Gemma, Llama, Mistral, Phi, Qwen) with LoRA via Unsloth, on spot GPUs with checkpoints to S3. - Produces standard outputs: a safetensors LoRA adapter, an optional GGUF file for llama.cpp/Ollama, and a model card. - A comparison report scores the fine-tuned model and the current LLM on the same examples and ends with a verdict: ready to replace, close, or not ready. - Works with coding agents through an MCP server (`uvx synqtune-mcp`) using read-only or launch-capable API tokens. ## Pages - [Home](https://www.synqtune.com/): what SynqTune does and how it works - [Use cases](https://www.synqtune.com/use-cases): tasks where a small fine-tuned model can replace an LLM API - [Pricing](https://www.synqtune.com/pricing): compute is billed by AWS to the customer's own account; savings calculator - [Documentation](https://www.synqtune.com/docs): setup, pipeline steps, troubleshooting, MCP server - [Security and AWS permissions](https://www.synqtune.com/security): every permission the IAM role has, what SynqTune does with it, what reaches SynqTune, and how to revoke it - [Changelog](https://www.synqtune.com/changelog): what changed, newest first ## Use cases - [Classification and routing](https://www.synqtune.com/use-cases/classification): One label per input, at the volume your queue actually has. - [Structured extraction](https://www.synqtune.com/use-cases/extraction): Documents in, JSON that matches your schema out. - [Tagging](https://www.synqtune.com/use-cases/tagging): Every tag that applies, zero or more per input. - [Redaction](https://www.synqtune.com/use-cases/redaction): Remove sensitive details, keep everything else exactly as it was. - [Summarization](https://www.synqtune.com/use-cases/summarization): Short, faithful summaries in the format your team reads. - [Regulated industries](https://www.synqtune.com/use-cases/regulated): Training data, compute and weights stay in your AWS account. ## Frequently asked questions ### Why would a small model be as good as a big one? Because it only does one narrow job. A frontier LLM is a generalist; a small model fine-tuned on thousands of your own examples becomes a specialist at your task, such as routing your tickets into your categories. It won't write essays, but on that one task it can match the big model, and the comparison report shows whether it does. ### Can I use my model in production today? Training and testing are ready today, free during early access. You can download the weights (a LoRA adapter, or a GGUF file for Ollama or llama.cpp) and run them yourself, or chat with the model on an instance in your account. A hosted endpoint your applications call directly is next on the roadmap, and early-access teams help shape it. ### How do I know the small model is good enough to switch? Add a comparison to the Evaluate step: your current LLM, with your production prompt, answers the same held-out examples as the fine-tuned model. The report scores both for your task, shows where the small model is weaker, and ends with ready to replace, close, or not ready. ### Why fine-tune a small model instead of using RAG or a frontier LLM? They solve different problems. A frontier LLM is best for open-ended work, and RAG for answering from documents that change. For a narrow, repeated task with examples of good answers, a fine-tuned small model learns the task itself, so each request carries only the input, costs GPU time instead of per-token fees, and stays in your AWS account. They also combine: retrieval can feed facts to your fine-tuned model. ### Does my data leave my AWS account? Training data and weights stay in your S3, and so do the evaluation examples. SynqTune stores pipeline settings, run status and evaluation scores and counts. If you turn on LLM grading or the comparison with your current LLM, records are sent to the provider you choose. ### Do I need ML or DevOps experience? No. Setup is one CloudFormation stack launched from the AWS console, and pipelines start from task templates with defaults. Every setting is visible and explained before you launch. ### What do I get at the end? A Hugging Face PEFT LoRA adapter (safetensors), a model card and README with SHA-256 checksums and evaluation results, and optionally a 4-bit Q4_K_M GGUF file for llama.cpp or Ollama. ### Which models can I fine-tune? Open models supported by Unsloth, typically from 1B to 8B parameters, such as Gemma, Llama, Mistral, Phi and Qwen. Templates pick a small default; you can enter any Hugging Face model ID under Advanced. ### What happens if a spot instance is interrupted? Checkpoints are written to your bucket during training. When the run is resumed it continues from the last checkpoint instead of starting over. ### How do I remove SynqTune's access? Delete the CloudFormation stack, or just the IAM role. SynqTune can no longer act in your account, and the S3 bucket and its contents are kept.