Measured open-model inference.
Priced by the yield, not the hype.

Inference Yield serves high-demand open-weight models on dedicated GPUs, with every serving configuration chosen by continuous benchmarking — quality-gated quantization, prefix caching with honest cached-token pricing, and published performance numbers measured on the live endpoint, never projected.

Current model

OpenAI-compatible API · streaming · tool calling · structured outputs · usage accounting on every request (cached tokens itemized).

How we're different

• Fastest visible endpoint for this model, measured: 162 tok/s per-stream single-user, 91 tok/s at 8-way concurrency — vs. 101 tok/s for the fastest marketplace rival as of our latest hourly market scan. Draft-model speculative decoding that passed our quality gate with zero output delta. Numbers from the live endpoint and live market data, never projected.
• Configuration is searched, not guessed. An automated tuner benchmarks runtime × precision × batching × speculative decoding for each model/GPU pair and promotes only configs that pass a quality gate (tool-calling, code, instruction-following, long-context) within 5pp of the bf16 reference.
• Prefix caching with real cached-token pricing. Repeated context bills at $0.08/M, itemized in usage.prompt_tokens_details.
• No content retention. Prompts and completions are never written to disk. Metadata-only metering, enforced by an automated test in our CI. Details.

Get started

curl https://inferenceyield.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen/qwen3.8-27b", "stream": true,
       "messages": [{"role": "user", "content": "Hello!"}]}'

Full examples in the docs. For keys, higher limits, or design-partner pricing: [email protected].