Service

Enterprise Inference API

Fast inference for compact LLMs on NVIDIA RTX PRO 6000 Blackwell GPUs — served through an OpenAI-compatible API.

01Features

Built for Speed

We focus on compact, efficient LLMs and serve each one from a single GPU — keeping the request path short for fast, responsive inference. Streaming output lets your users see tokens as they are generated.

One Model, One GPU

In the standard API, every model instance runs entirely within the 96 GB memory of a single GPU. No tensor splitting and no inter-GPU communication on the request path — for efficient, low-overhead inference.

OpenAI-Compatible API

Keep the SDKs and tools you already use. Point them to the Kobayashi endpoint, add your API key, and run your existing Chat Completions code with minimal changes.

Private by Design

Your prompts and outputs are never used to train models and are not retained after processing — only usage metadata is kept for billing. Inference runs entirely in Kobayashi data centers and is never forwarded to third-party model providers.

Secure API Key Access

All traffic is encrypted with TLS, and every request is authenticated with an API key. Create multiple keys for your teams and applications, and rotate or revoke them at any time.

Streaming, Tools & Vision

Build real applications, not just chat: streaming responses (SSE), tool/function calling, and image input are supported on compatible models.

02Hardware

Hardware & Models

NVIDIA RTX PRO 6000 Blackwell Server Edition

NVIDIA Blackwell architecture
96 GB GDDR7 GPU memory
1.6 TB/s memory bandwidth
5th-generation Tensor Cores with FP4 support
PCIe Gen 5 interface

Qwen3.8-27B

Default

Default model of the Enterprise Inference API
Open-weight vision-language model (Apache 2.0)
Served entirely on a single GPU
Also available on request: other open-weight LLMs that fit on a single 96 GB GPU, in BF16, FP8, or FP4 precision depending on the model

03Architecture

One Model per GPU, by Design

Our RTX PRO 6000 GPUs connect over PCIe. Instead of splitting a model across several GPUs — where every generated token waits on inter-GPU communication — we run each model instance entirely on one 96 GB GPU.

No tensor parallelism and no cross-GPU traffic on the request path
Model weights and KV cache held in a single GPU's memory
Multi-GPU tensor-parallel deployments available as custom projects

Standard API deployment

GPU
Full model

···

GPU
Full model

Each GPU serves a complete model

04API

OpenAI-Compatible Chat Completions

Already using the OpenAI Chat Completions API? Switch to Kobayashi by changing your base URL, API key, and model name. We support:

OpenAI-compatible Chat Completions
(works with the official OpenAI SDKs and compatible tools)
Streaming responses (SSE)
(show tokens to users as they are generated)
Tool calling & image input
(on supported models, including Qwen3.8-27B)

Python

from openai import OpenAI

client = OpenAI(
    base_url="https://<kobayashi-api-endpoint>/v1",
    api_key="YOUR_API_KEY",
)

stream = client.chat.completions.create(
    model="<model-id>",
    messages=[{"role": "user", "content": "Hello!"}],
    stream=True,
)

for chunk in stream:
    if chunk.choices:
        print(chunk.choices[0].delta.content or "",
              end="", flush=True)
print()

05Compare

Standard API and Custom Deployments

	Standard API	Custom deployment
Access	Shared API endpoint	Scoped per project
GPUs per model instance	1	1 or more
Multi-GPU tensor parallelism	No	Available
Models	Qwen3.8-27B; other single-GPU open-weight models on request	Scoped per project
API format	OpenAI-compatible (Chat Completions)	OpenAI-compatible (Chat Completions)
Authentication	API key	API key
Pricing	Contact us	Contact us

Inference performance varies with model, precision, context length, and load.

NVIDIA, RTX, and NVIDIA Blackwell are trademarks and/or registered trademarks of NVIDIA Corporation in the U.S. and other countries.

OpenAI is a trademark of OpenAI. “OpenAI-compatible” refers to API format compatibility only; Kobayashi Cloud is not affiliated with or endorsed by OpenAI.

Qwen models are developed by the Qwen team, Alibaba Cloud; Kobayashi Cloud is not affiliated with Alibaba Cloud. Qwen3.8-27B is released under the Apache 2.0 license.