Point at the checkpoint.
Bring an open weight or a LoRA. Name the GPU class you already rent. Dielet records a slice of your real traffic.
Open-model inference
Cut the die. Keep the tokens.
Point Dielet at an open checkpoint and the GPUs you already rent. Agents search kernels, batching, and decode layouts against recorded traffic, then serve the stack that measured faster — same silicon, identical outputs.
- Same GPUs measured on your mix
- Match first required before cutover
- Open weights Llama, DeepSeek, Mistral
Llama 3.1 70B
Llama 3.1 8B
DeepSeek V3
DeepSeek R1
Mistral Large
Open weights
Same GPUs
Llama 3.1 70B
Llama 3.1 8B
DeepSeek V3
DeepSeek R1
Mistral Large
Open weights
Same GPUs
Method
How Dielet ships
Point at a checkpoint. Let the mill search overnight. Swap the base URL when the die is cut.
Let the mill run.
Agents search kernels, batching, and decode layouts against that mix. Every candidate is measured, not guessed.
Cut over.
Swap the base URL. Tokens keep their shape. Latency drops on the same silicon. Serverless or dedicated.
Stack
The mill, in four cuts
Performance engineers are agents. You buy the endpoint those agents verified on your mix.
Profile
Agents sit on your GPUs and flag the true bottleneck: kernel, batch, decode, or HBM.
Search
Candidates include fused kernels, FP8 passes, and new shard maps. Each is measured.
Verify
A candidate ships only if outputs match and the clock is lower on your mix.
Serve
OpenAI-compatible endpoints. Dedicated capacity in under a day.
Keep
What does not change.
Your GPUs.
Dielet measures on the class you already rent. No new silicon required for a first cutover.
Your tokens.
Cutover waits until outputs match. The mill is not allowed to ship a faster wrong answer.
Your client.
Keep the chat client. Point it at a Dielet base URL after the mill prints a verified die.
Start
A die on the rack this week.
Write ceo@dielet.website. You get a test token, a hosted Turbo endpoint, and a 30-minute walkthrough on the model you want to serve.