Models/ Moonshot AI

Kimi K3.1 Pro, 2.8T parameters that read images

Moonshot AI's Kimi K3.1 Pro is an open-weight multimodal mixture-of-experts with 2.8T parameters, 104B active per token, post-trained for agent work. This endpoint serves it through chat completions.

  • 2.8T MoE, 104B active
  • Image input
  • 256K context
  • Open weights

Three tiers, per token.

Per 1M tokens

TierInputCachedOutput
Sync$0.24$0.20$8.402

openrelay/kimi-k3.1-pro

Full catalog on the inference pricing page. Add a card and make a deposit to start.

tools.py · function callingpython
from openai import OpenAI

client = OpenAI(
    base_url="https://inference.openrelay.inc/v1",
    api_key="or_••••••••",          # same SDK, new base URL
)

resp = client.chat.completions.create(
    model="openrelay/kimi-k3.1-pro",
    messages=[{"role": "user", "content": "What is the weather in Lisbon?"}],
    tools=[{"type": "function", "function": {
        "name": "get_weather",
        "parameters": {"type": "object",
                       "properties": {"city": {"type": "string"}}},
    }}],
)
print(resp.choices[0].message.tool_calls)

The spec sheet.

What the model is, what it takes in, and the surface it serves on.

Model id
openrelay/kimi-k3.1-pro
Parameters
2.8T total (mixture-of-experts), 104B active per token
Architecture
896 routed experts with 16 active per token plus 2 shared; 93 layers, 69 of Kimi Delta Attention and 24 of Gated DeepSeek Sparse Attention
Checkpoint
Moonshot's MXFP4 weights from quantization-aware training, served on AMD MI355X GPUs
Modality
Text and image in, text out
Context
256K tokens (262,144) on this endpoint, prompt and completion combined. The model supports up to 1M.
Endpoint
POST /v1/chat/completions, with tool calling and reasoning
License
Kimi K3.1 Pro License, open weights

Where Kimi K3.1 Pro earns its place.

What this model is actually for, versus the rest of the catalog.

Post-trained for agents

Moonshot trained K3.1 Pro from Kimi K3 with reinforcement learning across hundreds of domains, and reports gains in coding, financial research and factual accuracy.

Images on the same model id

A 401M-parameter MoonViT-V2 encoder reads images, so they go in as message parts on openrelay/kimi-k3.1-pro.

Trained in four-bit

Moonshot trained the model with MXFP4 weights in the loop, so the checkpoint we serve is the one it was trained as, not a quantization made afterwards.

Kimi K3.1 Pro, answered.

Is there a Kimi K3.1 Pro API?

Yes: openrelay/kimi-k3.1-pro, on the OpenAI-compatible chat completions endpoint at inference.openrelay.inc/v1. The OpenAI SDK works with a new base URL and an OpenRelay key.

What does Kimi K3.1 Pro cost?

The rate on this page is read from the catalog when the page renders. Input and output tokens bill separately, a cached prompt prefix bills at a lower input rate, and reasoning counts as output.

How much context can I send?

262,144 tokens (256K) on this endpoint, prompt and completion combined. Moonshot trained K3.1 Pro for up to 1M; this deployment runs a shorter window.

What hardware serves it?

AMD MI355X GPUs. One copy of the model spans the eight GPUs of a node, with a DFlash draft model for speculative decoding that halves the time per token on a single request.

First request in five lines.

Point the OpenAI SDK at inference.openrelay.inc/v1 and run Kimi K3.1 Pro per token. No contract, no minimums.