Modal launches Dojo, a library for production-grade LLM training

What Modal Dojo offers

Modal says reinforcement learning is hard even when infrastructure is solved. Teams still need appropriate base models, high-fidelity environments, clear and robust reward functions, and training recipes that work. Modal argues that a team with an agent harness in production, plus evals and agent traces collected from it, is already "80% of the way there". What is missing, in its view, is an easy way to start training on that data.

Dojo is meant to fill that gap. Modal describes the intended workflow as a single prompt: use Modal Dojo to train a chosen model on a proprietary dataset to replace a given agent.

The library builds on open-source frameworks, namely Miles and Slime. Modal says it has also patched and upstreamed a few pull requests itself. The company's pitch is that owning the full stack, down to the source code, lets teams tune recipes, build complex environments, patch the underlying frameworks and even reshape the observability dashboard.

Model, data and recipe

Modal's example shows a training run in under 100 lines of code. The user defines a model, a reward function, a dataset and a recipe, then launches the run through a config object. The example trains a Qwen model on a GSM8K dataset from Hugging Face with a custom reward function that checks the numeric answer.

The workflow has three parts:

• Model: Modal says the library has been validated on models from a half billion to multi-trillion parameters, with support for bringing your own. Teams run evals to find the right fit.
• Data: Users can load a dataset from Hugging Face, ingest a Harbor environment, or build one from their own harness and evals. Production traces can be used to generate tasks for the environment and form training and test splits.
• Recipe: The recipe defines how training happens, with parameters for hardware, parallelism, total rollout size, environment and more. Modal provides auto-optimized recipes for all supported models, and extending them for a specific use case means changing a few parameters, according to the company.

Observability and agent support

Dojo ships with an observability dashboard backed by a metric collector Modal built. It is meant to help catch reward hacking, debug stalling runs and track many concurrent runs. The dashboard shows loss and reward curves, framework-emitted metrics, the agent's trajectory for each rollout during training, profiling down to the substep, and container logs.

Modal also designed the library to be used by coding agents. Through the CLI, agents can fetch a run's status, metadata, logs and agent traces. The company says it has crafted skills that guide agents in a predictable way, so that an agent can "one-shot a baseline" for an RL problem. For more advanced setups, Modal invites users to get in touch.

Early customers and what it means

Modal cites two customers. Axiom Bio trains large open models to identify molecules from mass spectrometry data. Nishkrit Desai, Founding Engineer at Axiom Bio, said the team scaled up to 64k-context DeepSeek V4.1 Flash runs on 64 B300s and, within days, pushed performance further from the default recipes.

Afterquery, which curates expert data across domains, uses Dojo to train models like Kimi and DeepSeek at 256k context length. CEO Spencer Mateega said Dojo recipes enable autoresearch, so the research team can spend more time designing experiments.

Modal places Dojo after its earlier work on LLM inference with Endpoints, Endpoint Candidates for testing alternative deployments, and speculative decoding for trillion-parameter models. For teams that already run agents in production, the release offers a route from collected traces and evals to training open models on their own data, on infrastructure where they can still modify the underlying code. The source gives no pricing details.