Crusoe brings receipts to AI fine-tuning

Most fine-tuning services ask you to trust three things you can't verify: their benchmark harness, their headline numbers, and the bill that shows up after. 

Crusoe is taking the opposite bet: Publishing all three and letting customers check the math themselves. The company's Serverless Fine-Tuning is now generally available, allowing customers to fine-tune models with their own datasets more seamlessly. As part of this launch, the company ran its own benchmarks to help drive the best results possible, allowing customers to reproduce these tests and further fine-tune their models.

As of today, the service covers 19 open models ranging from 2 billion parameters all the way to 750 billion. Each of these models is run through the same benchmark suite, and is shipped with the exact configurations that produced the results, allowing customers to reproduce the tests. 

The Crusoe engineering team and lead PM Piyush Kadam led the benchmarking, with help from Emmanuel Acheampong on the developer advocacy team. Every model ran through an identical evaluation harness before general availability.

"You pick a model because a leaderboard said it was good, then you point it at your actual task and it falls apart," said Michael Yen-Chi Ho, director of technical product marketing at Crusoe. "The model isn't weak, it was just never trained on your specific datasets to tackle those problems. Fine-tuning the model fixes that." 

The biggest model tested is GLM 5.2, a 753 billion-parameter model. Crusoe's Serverless Fine-Tuning substantially improved the model when tested on its benchmarking suite, with its scores increasing 19 percentage points on its intent classification benchmark and 61 percentage points on its multi-turn conversational benchmark. Performing that fine-tuning run only cost $180. 

These results highlight that Crusoe's Serverless Fine-Tuning helps teams seamlessly and affordably make their models more accurate with their own datasets. 

"Fine-tuning teaches output discipline, not new knowledge," said Ho. "That distinction should decide how a team spends its budget, and most teams never get to see it clearly enough to know the difference."

Building a fine-tuning mechanism requires heavy-lift, from the underlying GPUs to the orchestration layer. However, Crusoe built Serverless Fine-Tuning as a service to address that. With simple drag and drop, users can select a model, point to the datasets they desire it to train on, and with pre-set configurations, the process is set. 

A few things stand out in what Crusoe published alongside the launch:

• Cost is transparent with tiered pricing by parameter count. Training runs from 40 cents per million tokens under 16 billion parameters up to $10 above 300 billion parameters, with no sales call required.

• Small models frequently outperform big ones on scoped tasks. A fine-tuned Qwen3.5-2B outperformed Qwen3-235B on Banking77, a benchmark for online banking customer service queries, at roughly a fifteenth of the training cost.

• The suite doesn't hide its own regressions. Two models, Qwen3.5-9B and Gemma-4-31B, scored worse after fine-tuning on a theorem-proving benchmark, a result Emmanuel Acheampong's team flagged as a likely sign of pre-training contamination rather than hiding it.

• Fine-tuned models can be deployed straight to dedicated endpoints, via Self-Serve Deployments starting at $5.50 an hour on Nvidia's H100 chips.

"The standard pitch for managed fine-tuning is trust our harness, trust our numbers, trust the bill," Ho said. "We publish the numbers and all parameters, and build them into smart-defaults so customers can pick and choose what works best for them, while enjoying an easy to use platform that requires little to no maintenance to integrate into their model ops workflow."

Upload a dataset and fine-tune one of the 11 GA models in Crusoe Intelligence Foundry.