DeepSeek, Qwen, GLM. One API.

Models and pricing

Every model, priced per token.

Input, cached-input and output rates for each model, with its context window and capabilities. Copy a model ID, or try the model in the playground.

USD per 1 million tokens. Context is the window we serve.

Model	Context	Input	Cached input	Output	Playground
DeepSeek V4 FlashFP8 DeepSeek·deepseek-v4-flash-0731 ReasoningReasoning ToolsTool calling JSONJSON output	1M	$0.13	$0.03	$0.26	Try
DeepSeek V4.1 FlashFP8 DeepSeek·deepseek-v4.1-flash-1017 ReasoningReasoning VisionImage input ToolsTool calling JSONJSON output	1M	$0.07	$0.003	$0.49	Try
Gemma 4 26B A4B ITBF16 Google·gemma-4-26b-a4b-it ReasoningReasoning VisionImage input ToolsTool calling JSONJSON output	256K	$0.11	$0.05	$0.37	Try
Gemma 4 31B ITBF16 Google·gemma-4-31b-it ReasoningReasoning VisionImage input ToolsTool calling JSONJSON output	256K	$0.14	$0.05	$0.40	Try
GLM-5.3FP8 Z.ai·glm-5.3 ReasoningReasoning ToolsTool calling JSONJSON output	230K	$1.21	$0.22	$3.86	Try
GLM-5.3 FlashFP8 Z.ai·glm-5.3-flash ReasoningReasoning VisionImage input ToolsTool calling JSONJSON output	1M	$0.14	$0.03	$0.45	Try
Nemotron 3.5 Super VL 120B A12B NVFP4 NVIDIA·nemotron-3.5-super-vl-120b-a12b-nvfp4 ReasoningReasoning VisionImage input ToolsTool calling JSONJSON output	1M	$0.15	$0.02	$0.50	Try
Qwen 3.8 27B FP8 Alibaba·qwen-3.8-27b-fp8 ReasoningReasoning VisionImage input ToolsTool calling JSONJSON output	256K	$0.30	$0.08	$2.35	Try

Input is the context you send. Output is what the model generates, including reasoning tokens. Cached input applies when the start of a prompt matches one the model has already processed. Sign in for current availability and prices in your billing currency.

What the API supports

The Chat Completions features you already use.

Each model’s row in the table shows which of these it supports.

• Streaming

  Set stream to true to receive tokens as server-sent events. Add stream_options to get token usage in the final chunk.

• Tool calling

  Models marked Tools can call functions you define. Your app runs the tool and sends back the result.

• JSON output

  Models marked JSON can follow a JSON Schema you provide, so replies have the shape your code expects.

• Reasoning controls

  Models marked Reasoning think before they answer. Turn thinking on or off per request, and set the effort where a model offers levels.

• Image input

  Models marked Vision accept images alongside text. Some also accept video.

• Long context

  Context windows up to 1M tokens. Each row in the table lists the window we serve.

Data and billing

Your prompts stay yours.

We serve each request on our own GPUs and keep no copy of it, and usage bills per token from the credit balance you already have.

• Prompts aren’t stored

  We don’t store your prompts or completions, and we don’t train on them.

• Cached input is automatic

  When the start of a prompt matches one the model has already processed, those input tokens bill at the cached-input rate. There is no cache-write fee and nothing to configure.

• One credit balance

  Token usage draws from the same Jarvislabs credit balance as your GPU instances. Add credits once and use both.

Get started

From first prompt to production.

Try a model in the playground, create an API key, then point your OpenAI client at Jarvislabs.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://models.jarvislabs.net/v1",
    api_key=os.environ["JARVISLABS_API_KEY"],
)

response = client.chat.completions.create(
    model="glm-5.3",
    messages=[
        {"role": "user", "content": "Say hello in one sentence."},
    ],
)
print(response.choices[0].message.content)

Inference on Jarvislabs

Pay per token, or run your own deployment.

Start on Token Factory, then move to a deployment of your own when you need dedicated capacity or a custom model.

Before you start

Common questions.

Practical details for choosing and using this product.