Product Overview

Two approaches, vastly different token costs. On the left, a coding query is sent directly to Claude. On the right, the same query is routed to a smaller, cheaper model, more appropriate for the query's complexity. The result is stark. This model just routed that coding request to a model a hundred times smaller. Same answer, ten times cheaper, running on hardware you already own. Routing everything through one big rented model is the expensive default. Every AI request you send to a Frontier API costs metered tokens and you break the feedback loop. But this doesn't have to be. Fine tuning platforms give you the option to ditch the Frontier for everything and use inference more intelligently. You still do the engineer's job: evals, test sets, data curation, failure analysis and iteration. Oumi automates that whole workflow. Think Claude code but for AI engineering. And in this video, I'll show you how. We'll build a coding model router from a single prompt. No data required. Let's log in to the Oumi platform. You can chat with Oumi just like you would chat with Claude Code. Only Oumi is specialized for AI development, and we call this Agentic AI Engineering. Let's enter our prompt, and the prompt is as follows. I want to build a model router specifically for deciding between coding models. Given a coding question, the router must analyze its complexity, category, and requirements, then recommend the most cost effective open model. All called Sonnet and Opus for the most complex tasks to answer it. You must balance quality against cost. Use cheaper, smaller models for simple tasks and reserve expensive, larger models for complex ones. So after we've entered the prompt, the Oumi agent is now going to ask clarifying questions, in this case about the output format. We'll choose JSON and specify two fields, model chosen and the rationale, and it also asks about which models the router needs to decide between. Now at this stage, the Oumi agent decides that it has enough information about the task, and it devises a plan. And the plan is as follows. So define the metric, synthesize test data, evaluate the baseline models, synthesize training data, then fine tune the model, and evaluate the model after fine tuning. Finally, doing a review and deciding on next steps. So you can see that the different steps ask additional clarifying questions and launch jobs on Oumi managed infrastructure. It walks us through the entire process of model development. We've chosen Qwen three point five 4B as the base model, the one being fine tuned, and OPUS four point six as the model that synthesizes the training data. However, the platform supports a comprehensive range of open and closed weight models for fine tuning data synthesis and evaluation by LLM judges. After training, the evaluation metrics showed that we were able to, from our fine tuning, drastically improve accuracy and other metrics over the small base model, and also match OPUS four point six performance, with a model that is 100x smaller and 10x cheaper to run. In the example in this video, we distilled OPUS four point six into a small model, and so we expect the student, our base model, to match the teacher OPUS four point six. However, when fine tuning to external data, the fine tuned model typically outperforms the frontier model by ten percent or more. There's an example of that right here. Check out the rent versus own calculator at oumi dot ai slash calculator to get an estimate of how much you can save by building your own custom AI for your specific use case. So we've turned requiring metered tokens from a Frontier model forever to a fixed cost of a few dollars to train a small language model. But we can do even better. Both before the model is put into production and afterwards, valuable feedback signals exist, if only they could be captured and utilized. Fortunately, Oumi has several features for doing just that, and we call this the self improvement loop. Let's examine the failure modes of the fine tuned model. What this means is the high level groupings of the types of prediction errors observed. Oumi takes those failure modes and curates a training set targeted at exactly those gaps. Oumi can continuously improve your models, performing the same steps as we've just shown after the model has been put into production to learn from failed predictions, detect distribution drift and more. Stop throwing away valuable learning signals when you build with Anthropic or OpenAI and start owning the self improvement loop. Don't want to handle the inference infrastructure? That's okay. Oumi can handle it for you with one click deployments. And to use the model in your existing code, it's as simple as swapping out your endpoint address for Oumi's OpenAI compatible endpoint. A single line change. On the other hand, if you do need to fully manage inference, you have complete control over the model weights. You can deploy to your on premise infrastructure or even directly to consumer devices. Oumi gives you model sovereignty and full deployment options, whether that's Oumi managed in your Cloud VPC, on premise or on edge devices. Here's an example. You can take your Oumi trained model, make the model even smaller by applying quantization, and then deploy it to a consumer's smartphone. You brought a task and Oumi synthesized, evaluated, trained and deployed. Minutes of human time, and now you own the results. We've built a calculator for your AI tasks' total cost of ownership. Describe your use case, and we'll estimate costs and start building a prototype just for you. You can book a time to meet with an Oumi AI expert and see your AI prototype in action. Until then, I hope to see you on the platform. We're offering free credits for new users, and may the AI vibes be with you.