Make your inference engine faster, or you don't pay
- 1 Time
- 2 Setup
- 3 Confirm
November 2026
MoTuWeThFrSaSu
Pick a day to see times.
I make inference engines faster on your hardware: open-source engines like llama.cpp, vLLM and TensorRT-LLM, or your own closed-source one. You get a fixed quote with a speed target written in. If the target is not met, you pay nothing.
How it works
- You tell me the setup: engine, model, GPU, and the workload you care about (one user or many, short or long context).
- I measure it first, free. BoltBeam audits your setup against the roofline: the fastest your GPU can run that model. The gap is what can be won, and it tells both of us whether the job is worth doing.
- You get a fixed quote with one test in it. For example: decode at least 15% faster at batch 1 on your workload, measured by BoltBeam, before and after, paired runs on the same machine.
- I do the work. Engine patches, kernels, settings, or a lighter checkpoint, whichever closes the gap.
- The test decides. Same benchmark, same machine. If it passes, the quote is charged. If it does not, nothing is.
The guarantee
- Card saved at $0 when you book. Nothing is charged for the audit.
- Target met: the quoted amount is charged.
- Target missed: nothing is charged, and you keep the audit.
Open or closed source
Open-source engines: the work can go upstream or stay as patches you own, your choice. Closed-source engines: fine, with your written permission to profile and modify the engine, under an NDA. The work runs on your hardware or a rented machine you approve, and the results stay private unless you want them published.
What I have shipped
- llama.cpp v1.0.0 at 2.92x stock on a 128k-token session, RTX 5090, with no measured quality loss.
- A 4-bit output head for vLLM checkpoints: +10% on Qwen3.8-27B, 98.1% the same top token.
- BoltBeam, the open-source tool every number on this site comes from.