DeepInfra raises $107M Series B to scale the inference cloud — read the announcement

Kimi K3: 2.8T Open-Weight Multimodal Model

Published on 2026.08.11 by DeepInfra
Kimi K3: 2.8T Open-Weight Multimodal Model

Updated November 2026.

Kimi K3, developed by Moonshot AI, represents a landmark achievement in open-source artificial intelligence. As a 2.8-trillion-parameter native multimodal Mixture-of-Experts (MoE) model, Kimi K3 is engineered to handle demanding computational tasks, from complex software engineering and long-horizon agentic workflows to deep scientific research.

By combining a one-million-token context window with its architectural innovations, Kimi K3 offers frontier-level performance that rivals the world’s leading proprietary models. This overview covers the model’s capabilities, performance benchmarks, and how to integrate it via the DeepInfra platform.

Moonshot has since released Kimi K3.1 Pro, post-trained from Kimi K3 with sparse attention for long context. It is available on DeepInfra at $1.90 input and $9.50 output per 1M tokens (current promotional pricing), with a Flex tier. See the Kimi K3.1 Pro model overview.

Architectural Innovation and Efficiency

Kimi K3 introduces several breakthroughs designed to maximize reasoning power while maintaining operational efficiency. Unlike traditional dense models, Kimi K3 utilizes a Stable LatentMoE framework. While the model contains 896 experts, it activates only 16 per token, totaling 104 billion active parameters. This design results in a 2.5x improvement in scaling efficiency compared to its predecessor, Kimi K2.

Key Technologies

  • Kimi Delta Attention (KDA) & Attention Residuals (AttnRes): These mechanisms optimize long-context processing, enabling the model to maintain high performance across its entire 1M-token window.
  • Native Multimodality: Kimi K3 is built from the ground up to process text, images, and video within a single unified framework, rather than relying on external plugins.
  • Quantization-Aware Training (QAT): The model features native MXFP4 weights and MXFP8 activations, ensuring broad hardware compatibility and efficient inference without sacrificing accuracy.

Performance Benchmarks

Kimi K3 consistently delivers top-tier results across reasoning, coding, and vision benchmarks, often setting the standard for open-weight models and matching proprietary systems like GPT-5.6 Sol and Claude Fable 5. The figures below are vendor-reported by Moonshot at launch in July 2026.

Comparative Performance Overview

CategoryBenchmarkKimi K3 (Max)Claude Fable 5GPT-5.6 Sol
ReasoningGPQA Diamond93.592.694.1
CodingSWE-Marathon42.035.039.0
AgenticBrowseComp91.288.090.4
VisionVideo-MME (w/ sub)90.0—89.5
VisionMMMU-Pro81.6 / 83.481.2 / 86.583.0 / 84.6

Specialized Capabilities

  • Engineering & Coding: Kimi K3 is optimized for long-horizon tasks, sustaining engineering sessions across large repositories and orchestrating terminal tools for GPU kernel optimization and compiler development, with minimal human oversight.
  • Agentic Knowledge Work: The model advances end-to-end knowledge work, producing deep research with interactive visualizations, widgets, and dashboards. It scored 94.5 on MCPMark-Verified, indicating strong reliability in tool-calling environments.
  • Document & Video Intelligence: With a score of 91.1 on OmniDocBench, Kimi K3 leads in document layout analysis, and its native multimodal architecture also supports motion design and video editing tasks.

Getting Started with the Kimi K3 API

DeepInfra provides an OpenAI-compatible interface for Kimi K3, making it easy to integrate into existing workflows. The model is listed on the DeepInfra model catalog alongside DeepInfra’s broader lineup of open-weight models.

Authentication

To interact with the API, you must use a DeepInfra API key. Retrieve your key from your DeepInfra Dashboard and pass it as a Bearer token in your HTTP headers:

Authorization: Bearer <YOUR_DEEPINFRA_API_KEY>

API Endpoint Basics

  • Base URL: https://api.deepinfra.com/v1/openai
  • Chat Completions Path: /chat/completions
  • Method: POST

Using cURL

curl https://api.deepinfra.com/v1/openai/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $DEEPINFRA_API_KEY" \
  -d '{
    "model": "moonshotai/Kimi-K3",
    "messages": [
      {
        "role": "user",
        "content": "Explain the benefits of Kimi Delta Attention for long-context reasoning."
      }
    ],
    "temperature": 0.2
  }'
copy

Using Python

import os
import requests

api_key = os.getenv("DEEPINFRA_API_KEY")
url = "https://api.deepinfra.com/v1/openai/chat/completions"

headers = {
    "Authorization": f"Bearer {api_key}",
    "Content-Type": "application/json"
}

payload = {
    "model": "moonshotai/Kimi-K3",
    "messages": [{"role": "user", "content": "Write a Python function to optimize a GPU kernel."}],
    "temperature": 0.2
}

response = requests.post(url, headers=headers, json=payload)
print(response.json())
copy

Configuration and Parameters

When sending requests to the /chat/completions endpoint, you can fine-tune the model’s behavior using the following parameters. The full parameter reference is available in the Kimi K3 API documentation.

ParameterTypeRequiredDescription
modelstringYesUse moonshotai/Kimi-K3.
messagesarrayYesSupports text and image inputs for multimodal tasks.
temperaturenumberNoControls randomness (e.g., 0.2 for deterministic outputs).
response_formatobjectNoUse {“type”: “json_object”} for structured data.
toolsarrayNoDefine functions for agentic tool-calling workflows.
max_tokensintegerNoMaximum tokens to generate, up to 1,048,576. If unset, DeepInfra uses the smaller of 65,536 and the remaining context window. There is no 16,384-token output cap on this model.
reasoning_effortstringNonone, minimal, low, medium, high, xhigh or max. Set “none” to turn thinking off. Reasoning tokens are billed as output tokens.

Pricing and Token Usage

DeepInfra offers a transparent, usage-based pricing model for Kimi K3. The model supports cached input pricing, which significantly reduces costs for repeated long-context tasks by billing cached tokens at a fraction of the standard input rate.

Token TypePrice per 1M Tokens
Input Tokens$2.85
Output Tokens$14.25
Cached Input Tokens$0.285
Prompt cache write, 5-minute retention (1.25×)$3.5625
Prompt cache write, 1-hour retention (2×)$5.70

Pricing as of 2026-11-17. Kimi K3 has no promotional rate and no Flex or Priority tier on DeepInfra. These rates are 5% below Moonshot’s own $3.00 input, $15.00 output and $0.30 cached input per 1M tokens.

For long agent sessions, you can keep a prompt prefix cached for 5 minutes or 1 hour. Cache writes are retained in whole blocks of 1,536 tokens (the remainder is billed as standard input), and reuse within the window bills at the $0.285 cached rate. See prompt cache retention for details, and the DeepInfra pricing page for current rates.

Final Thoughts

Kimi K3 stands as a powerful, cost-effective alternative to proprietary frontier models. Its combination of a 2.8T-parameter architecture, 1M-token context window, and native multimodality makes it a strong choice for developers building the next generation of AI agents and complex engineering tools. Here are the key takeaways:

  • Frontier Performance: Matches or approaches top-tier proprietary models in coding and reasoning.
  • Architectural Efficiency: 2.5x scaling improvement via KDA and MoE.
  • Developer Friendly: OpenAI-compatible API with support for JSON mode and function calling.
  • Newer Option: Kimi K3.1 Pro, post-trained from Kimi K3, costs about a third less on DeepInfra at current promotional pricing and adds a Flex tier.

To begin using Kimi K3 for your projects, visit the DeepInfra Dashboard to retrieve your API key or deploy a private endpoint for dedicated capacity.

Related articles
GLM-5 API Benchmarks: Latency, Throughput & CostGLM-5 API Benchmarks: Latency, Throughput & Cost<p>Updated November 2026. GLM-5 has been retired on DeepInfra. Its successor, GLM-5.3, is available now; see the GLM-5.3 API provider benchmarks for current figures. The benchmark analysis below is kept for reference. GLM-5 is the latest open-weights reasoning model released by Z AI (Zhipu AI) in February 2026, characterized by high &#8220;thinking token&#8221; usage. It [&hellip;]</p>
TypeSafe Jev Review: Testing 5 Jobs on a Non-Chat ModelTypeSafe Jev Review: Testing 5 Jobs on a Non-Chat Model<p>Jev is an AI model from TypeSafe that makes decisions instead of writing text. A chatbot like ChatGPT reads your prompt and generates an answer word by word. Jev reads a piece of text, such as a support ticket or a request from an AI agent, plus questions you&#8217;ve written about it, and hands back [&hellip;]</p>
MCP Servers Explained: A Practical Guide for DevelopersMCP Servers Explained: A Practical Guide for Developers<p>Giving an AI application access to a database, API, file system, or internal service sounds simple until you have to maintain the integration. You need to define what the model can access, describe operations, execute requests safely, return structured results, and handle authentication. You may need to repeat much of that work when another AI [&hellip;]</p>