So, like, Together.ai just dropped this new thing called the Together Code Interpreter (TCI). It’s basically an API that lets developers run code generated by Large Language Models (LLMs) with ease. This is supposed to make life easier for peeps using LLMs for coding and stuff, according to the folks at Together.ai.
Making Code Execution a Breeze
LLMs are great at spitting out code, but they usually can’t actually run it, which means developers have to do all the testing and debugging themselves. TCI steps in to solve this problem by offering a simple way to safely execute LLM-generated code on a large scale. This is meant to make developing agentic workflows smoother and open up possibilities for more advanced reinforcement learning operations.
What’s the Deal with TCI?
So, the Together Code Interpreter works by taking code from LLMs, running it in a secure sandbox, and showing you the results. You can then feed this output back into the LLM for continuous improvement in a closed-loop system. This whole process allows for more dynamic responses from LLMs, which is pretty cool. One example they give is if an LLM like Qwen Coder 32B spits out some code for making a chart, TCI can run it and show you the visual result, bypassing the LLM’s usual limitations.
Boosting Reinforcement Learning
TCI’s quick code execution abilities have caught the eye of machine learning teams focusing on reinforcement learning. It lets you automate evaluation through thorough unit testing, making training cycles more efficient. With TCI, you can handle tons of sandbox executions at once, which is key for rigorous testing and evaluation. Berkeley AI Research and Sky Computing Lab’s open-source project Agentica has already hopped on the TCI train, speeding up their training cycles and improving model accuracy while keeping costs down.
The Nitty-Gritty Details
Together.ai has come up with this thing called “sessions” to measure TCI usage, priced at $0.03 per session. Each session lasts 60 minutes and supports multiple execution jobs, making it easy to scale TCI across different applications. To get started with TCI, devs can use the Python SDK or API provided by Together.ai, along with detailed documentation. They’ve also added support for MCP, so you can integrate code interpreting into any MCP client, expanding the tool’s reach and usefulness.
In a nutshell, Together Code Interpreter is changing the game for how developers handle LLM-generated code, offering a smooth, scalable solution for running complex workflows and boosting machine learning operations.
