Thermite for AI and machine learning
Faster inference and better models on the hardware you own.
Inference engines, LLM applications and training recipes each have a metric that defines success, and agents improve it on your own GPUs while guards hold output, cost and accuracy. A goal campaign fits a model to a memory and speed budget; a research campaign reads the papers and model cards behind a choice of method.
LLM inference optimization
Agents tune an engine's kernels and launch settings for your GPU. The judge times decode on that card and holds the model's outputs to a reference.
For Local AI · Startups
- Measures
- Decode tokens a second
- Holds
- Outputs held to a reference
- Hidden
- Prompts the agents never see
LLM application quality
Agents change prompts, examples, schemas, retrieval and the model for each step. The judge runs the pipeline on your labelled examples, scores it with a program, and guards cost and time.
For Startups · Enterprises · Local AI
- Measures
- Field F1 or accuracy on your labels
- Holds
- Cost and time per example held
- Hidden
- A test split the agents never see
Model training efficiency
Agents change the model, the optimizer and the schedule. The judge trains each version for the same fixed time on your GPU and scores it on validation text the agents never see.
For Local AI · Solo builders
- Measures
- Validation bits per byte after a fixed time
- Holds
- The time limit on the judge's clock
- Hidden
- Validation text the agents never see
Model and method landscape
Agents read the papers, model cards, benchmark repositories and issue threads on each model and method you're weighing. Code checks each quote, and a review panel of models decides each question.
For Local AI · Startups · Enterprises
- Counts
- Verified claims on each model and method
- Checked
- Each quote and number found in its archived page
- Decided
- A panel of models, question by question
Fitting a model to your hardware
Agents choose the weight file, the cache and the server's settings. The judge serves the model on your card, checks its memory, speed, quality and API, and the run ends when every goal passes twice.
For Local AI · Solo builders · Startups
- Measures
- Peak memory, decode speed and quality on your card
- Holds
- Every goal met on two checks
- Hidden
- Questions and a document the agents never see
Guides by audience
Twenty minutes to a first verdict.
Install it on one computer, or ask your own AI agent to:
Set up Thermite on this computer for me: follow https://usethermite.com/llm-setup.txt