Workflow · Production

Replit builds code completion model in under a week with Databricks Mosaic AI Training

The problem

Replit needed to train a specialized code completion model for their developer platform but lacked an efficient end-to-end training infrastructure, and all available alternatives were either underdeveloped or too complex for their small engineering team.

First attempt

Before choosing Databricks, Replit evaluated multiple alternative training platforms but found them all either behind in capabilities or exposing unmanageable complexity for a small team.

Workflow diagram · grounded in source
1
Code completion need identified
trigger
“Replit knew they needed to train their own specialized model for code completion in order to build this platform”
2
Alternatives evaluated and rejected
validation
“Before working with Databricks, Replit explored various options but found them to be either underdeveloped or overly complex”
3
Mosaic AI Training adopted
integration
“Replit leveraged the Mosaic AI Training infrastructure and tools to experiment with smaller models”
4
Model scaled to 256 GPUs
ai_action
“gradually scaling up to a larger allocation of 256 GPUs just a week before the software company launched its code completion feature”
5
LLM trained and code completion launched
output
“Replit successfully conducted a "YOLO" run of its LLM and launched its code completion model in time for its developer day”
6
Continuous optimization cycle
feedback_loop
“The journey from the initial version of the model to production demonstrated a continuous cycle of learning, adapting and optimizing”
Reported outcome

Replit built its code completion model from scratch in less than a week using Mosaic AI Training, launched on time for developer day, and massively increased the productivity of their AI engineers with faster time to market for new models.

Reported metrics
Time to build code completion modelless than a week
GPU allocation for model training256 GPUs
AI engineer productivitymassively increased
Time to market for new modelsfaster time to market
Show all 5 reported metrics
time to build code completion modelless than a week
GPU allocation for model training256 GPUs
AI engineer productivitymassively increased
time to market for new modelsfaster time to market
business impactsignificant business impact
Reported stack
Mosaic AI TrainingDatabricks
Source
https://www.databricks.com/customers/replit
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Replit built its code completion model from scratch in less than a week using Mosaic AI Training, launched on time for developer day, and massively increased the productivity of their AI engineers with faster time to…

What tools did this team use?

Mosaic AI Training, Databricks.

What results were reported?

Time to build code completion model: less than a week; GPU allocation for model training: 256 GPUs; AI engineer productivity: massively increased; Time to market for new models: faster time to market (source-reported, not independently verified).

What failed first in this deployment?

Before choosing Databricks, Replit evaluated multiple alternative training platforms but found them all either behind in capabilities or exposing unmanageable complexity for a small team.

How is this workflow AI workflow structured?

Code completion need identified → Alternatives evaluated and rejected → Mosaic AI Training adopted → Model scaled to 256 GPUs → LLM trained and code completion launched → Continuous optimization cycle.