Technical profile
A runtime built around the model.
Kernelwright applies a repeatable optimization discipline across model launches: profile the model and hardware together, then tune kernels, memory layout, quantization, KV cache, batching, and serving policy as one execution system. Performance is reported from exact-hardware, end-to-end measurement.
Interface
Compatibility contract
The serving interface is designed for marketplace routing and direct product integration, with model-specific capabilities declared explicitly.
| Capability | Commitment | Standard |
|---|---|---|
| Chat completions | OpenAI-compatible request and response shapes | Core contract |
| Streaming | Server-sent events with measured time to first token | Core contract |
| Tool calling | Compatible tool definitions and structured calls | Model-qualified |
| Structured output | Schema-constrained response support | Model-qualified |
| Usage | Prompt, completion, and total-token accounting | Core contract |
| Admission control | Early 429 responses instead of unbounded queues | Core contract |
Measurement standard
What Kernelwright measures
Aggregate prompt and completion throughput across all requests
Time to first token and inter-token latency by concurrency level
Correctness across text, streaming, tools, and structured output
Queue delay, rate limits, errors, retries, and recovery behavior
Sustained thermal and memory behavior on the exact host
Serving cost and contribution at the observed request mix
Capacity policy
No benchmark theater.
Published capacity will reflect sustained validated throughput with normal compatibility features enabled. If demand exceeds that capacity, the endpoint will reject early rather than hide latency in a long queue.