Model Evaluation: Codestral 25.08

Process outline

Process

Identify candidate model(s)

Codestral 25.08

Target feature

Code Completion

Data privacy

Hosted on Fireworks.ai, same as now

Context window

256K tokens

Global availability

All regions as it's hosted on Fireworks.ai

Fill-in-Middle (FIM) support

Yes

Language support

At least as much as Codestral 2501, our current model, however information isn't publicly available

Industry benchmarks

[Include any industry benchmarks]

Internal evaluation

codestral-2501-baseline-b6f73622_codestral-2508-c8f78ffe.csv

Latency

Note: Codestral 2508 was deployed on A100s instead of H100s and was FP16 precision instead of MM2 FP8. Latency isn't comparable in this case, Fireworks would help us optimize the new model before deploying to production

Quality

Load test

[Include the results of the load tests]

Other Notes

[Include any other relevant information here]

Internal rollout

Link to rollout issue

External rollout

Link to rollout issue