Produck Inference

Production LLM Inference

Use selected models through an OpenAI-compatible chat completions API. Model IDs, context windows, and input/output rates are listed for each model.

Explore models Get API key Read the docs

Example request

curl

curl -X POST \
  https://api.produckai.com/v1/chat/completions \
  -H "Authorization: Bearer $PRODUCK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-oss-120b","messages":[{"role":"user","content":"Hello"}]}'

Models

Compare models by price and context.

All models

ModelInput / 1MOutput / 1MContextAction

GPT-OSS-120BGPT-OSS-120B is available through Produck for text chat and streaming with a 131K context window.gpt-oss-120b

$0.04

$0.18

131K

View model

Gemma-4-31BGemma-4-31B is available through Produck for text chat and streaming with a 262K context window.gemma-4-31b

$0.14

$0.40

262K

View model

DeepSeek V4 FlashDeepSeek V4 Flash is the long-context option in the current Produck catalog, with a ~1M context window for text chat and streaming.deepseek-v4-flash

$0.09

$0.20

~1M

View model

Qwen3 Next 80BQwen3 Next 80B is available through Produck for text chat and streaming with a 262K context window.qwen3-next-80b

$0.10

$1.10

262K

View model

Build with HTTP

OpenAI-compatible chat completions API

Use the supported /v1/chat/completions endpoint for text chat and streaming. Send a Produck API key in a Bearer authorization header.

Start with the quickstart

POSThttps://api.produckai.com/v1/chat/completions

One endpoint, explicit model IDs, ordinary HTTP requests.

At a glance

Model and API details in one place.

Input and output rates

Token prices are listed separately for every model.

Model catalog

Compare model IDs, context windows, and prices in one table.

Inference engineering

Serving work starts with workload shape, latency goals, throughput, and reproducible evidence.

Engineering

Inference is a systems problem.

Characterize the workload, measure p95 TTFT and TPOT, inspect throughput, batching, KV-cache behavior, and GPU utilization, then validate changes against the target SLO. Our engineering work spans vLLM, SGLang, capacity planning, and dedicated deployment design.

Explore Inference Engineering

01 Measure the workload02 Identify the bottleneck03 Validate the result

Research

Inference systems research.

A controlled A100 study of vLLM concurrency, throughput, latency, GPU behavior, and the post-knee regime.

Explore research

Featured investigation

Beyond Peak Throughput: Finding and Explaining vLLM’s Concurrency Knee on an A100

Measured throughput and latency, repeated saturation runs, and a trace-based explanation of the post-knee regime.

Read investigation

Get started

Choose a model and make your first request.

Review the model catalog, create an API key, or bring us a specific inference workload.

Get API key Discuss a workload