Tune it here, serve it here
Because we own the serving stack, an adapter you train can go live behind the same model ID your application already calls — no new endpoint, no second SDK.
LoRA by default
Rank-16 to rank-128 adapters train in hours and load in seconds, so a tuned variant costs the same to serve as the base model.
Full fine-tune on request
For domain shifts a LoRA cannot cover, full-parameter runs on reserved capacity, priced per GPU-hour.
Same model ID
Serve the adapter as llmcloud/<model>:<tune>. Routing policies, caching and analytics keep working unchanged.
Customer-provided tunes first
Serving a compatible checkpoint on reserved capacity is available to design partners. Managed training on llmcloud remains a separate roadmap capability.
Eval before promote
Every run reports against your held-out set and the base model. Promote to production only when it wins.
Export the weights
Adapters and full checkpoints are yours to download and run elsewhere, including self-hosted.
POST /v1/fine-tunes
{
"base_model": "llmcloud/qwen3-235b-a22b",
"method": "lora",
"rank": 32,
"training_file": "file_9f2c...",
"eval_file": "file_1a7d...",
"region": "eu-west"
}{
"model": "llmcloud/qwen3-235b-a22b:support-v3",
"messages": [{"role":"user","content":"…"}]
}When does this ship?+
LoRA on the DeepSeek and Qwen families first, then Llama and gpt-oss. Waitlist members get access as each family opens.
What will it cost?+
Training is billed per GPU-hour at the dedicated-capacity rate. Serving a LoRA adapter costs the same per token as the base model — we do not surcharge tuned variants.
Can I tune a closed frontier model?+
No. Tuning is limited to open weights we host. For closed models, use the provider's own tuning API through BYOK.
Which regions?+
Wherever the base model already runs, including sovereign regions once those go live.