Now on VeniceLLMReasoningOpen weights

Kimi K3.1 Pro

Moonshot AI's 2.8T-parameter open-weight MoE flagship — 1M-token context, native vision, and frontier agentic coding in a fully private, zero-retention runtime.

For agents
curl https://api.venice.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "e2ee-kimi-k3-1-pro-p",
    "messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
  }'
Model IDe2ee-kimi-k3-1-pro-p
Maker
Moonshot AI
Context
1,000K tokens
Reasoning
Supported
Privacy
E2EE

Overview

What is Kimi K3.1 Pro

Kimi K3.1 Pro is Moonshot AI's most capable open-weight model: a 2.8T-parameter mixture-of-experts (104B active) with a 1M-token context window, native image understanding, tool calling, and web search. Post-trained from Kimi K3 with large-scale reinforcement learning, it excels at agentic coding, financial research, and long-document work.

Running it privately on Venice

On Venice, Kimi K3.1 Pro runs in a private, zero-retention environment: your prompts are never stored, profiled, or used for training, and the deployment is end-to-end encrypted inside a trusted execution environment. You get a frontier open-weights model without feeding a Big-Tech data pipeline — billed per token, with no subscription and no account history tied to your usage.

PrivateNo prompt trainingTEE · hardware enclaveEnd-to-end encrypted

Agent quickstart

Three calls, copied straight out

The API is OpenAI-compatible: change the base URL and the model id and existing client code works unchanged.

Streaming chat

curl https://api.venice.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "e2ee-kimi-k3-1-pro-p",
    "stream": true,
    "messages": [{ "role": "user", "content": "Draft the release note." }]
  }'

Tool calling

curl https://api.venice.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "e2ee-kimi-k3-1-pro-p",
    "messages": [{ "role": "user", "content": "Find the rate limits." }],
    "tools": [{
      "type": "function",
      "function": {
        "name": "search_docs",
        "description": "Search the API documentation.",
        "parameters": {
          "type": "object",
          "properties": { "query": { "type": "string" } },
          "required": ["query"]
        }
      }
    }],
    "tool_choice": "auto"
  }'

Python SDK

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["VENICE_API_KEY"],
    base_url="https://api.venice.ai/api/v1",
)

resp = client.chat.completions.create(
    model="e2ee-kimi-k3-1-pro-p",
    messages=[{"role": "user", "content": "Build without permission."}],
)
print(resp.choices[0].message.content)

Specifications

Datasheet

Maker
Moonshot AI
Modality
Text and image input, text output
Open weights
Yes
License
Kimi K3.1 Pro License (permissive; MaaS revenue clause above $20M/yr)
Modes
Selectable reasoning effort (Moonshot benchmarks at max); streaming and tool-calling via the API
Context window
1,000K tokens
Released
November 4, 2026 (weights); announced October 23, 2026
Architecture
Mixture-of-Experts transformer, 93 layers, KDA + Gated DSA attention, MoonViT-V2 vision encoder (401M params)
Parameters
2.8T total, 104B active per token, 896 experts (16 selected), 160K vocabulary
Quantization
MXFP4 weights / MXFP8 activations (quantization-aware training)
Max output
128K tokens
Capabilities
Vision, Function calling, Reasoning, Web search, Code-optimized
Privacy on Venice
Private — zero retention
Available on Venice since
Nov 2026

Assessment

Strengths and limitations

Strengths
  • Frontier agentic coding: Moonshot reports 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra, and Artificial Analysis ranked it 4th overall on its Intelligence Index — the top open-weights model.
  • Genuinely open weights under a permissive licence, so the same model you use on Venice can be self-hosted, fine-tuned, or audited.
  • 1M-token context window with 128K max output: whole codebases, long contracts, or months of logs fit in a single request.
  • Full agent stack on Venice: tool use / function calling, structured JSON output, web search, and multiple image inputs per request.
  • Natively multimodal via the MoonViT-V2 vision encoder — charts, screenshots, and documents can be reasoned over, not just transcribed.
  • Runs privately on Venice with zero retention, end-to-end encryption, and TEE isolation — sovereignty over your prompts and code.
Limitations
  • Not uncensored: Venice applies its content policies to this model; if you need an uncensored open model, Venice Uncensored 1.2 is the option on the platform.
  • Output tokens cost $12.50/1M: roughly 3–6× rivals like Qwen 3.8 Max or Gemini 3.8 Flash, and agentic runs burn many tokens.
  • It is a very large model (2.8T total parameters); third-party testers report long runtimes on complex multi-step tasks, and Venice does not currently list the faster K3.1 Pro Highspeed variant.
  • The licence's Model-as-a-Service clause requires a separate agreement with Moonshot AI for providers exceeding $20M in annual revenue — relevant if you plan to resell inference.
  • Some headline benchmarks (Kimi Code Bench 3.0) are Moonshot's own internal tests, so treat vendor-reported numbers with normal skepticism.

Use cases

What it is good for

  1. 01Agentic coding: multi-step bug fixing, refactors, and terminal-driven engineering tasks across large repositories.
  2. 02Financial research: ingesting long filings, earnings transcripts, and tables within the 1M-token window.
  3. 03Document question answering over lengthy PDFs, contracts, and mixed text-and-image reports.
  4. 04Retrieval-augmented generation pipelines that need tool calling and structured JSON output.
  5. 05CAD scripting and 3D-creation workflows, two areas Moonshot specifically targeted in post-training.

Prompting

Getting better results

Paste the entire repo, log, or filing — with a 1M-token context there is no need to summarize first; ask it to cite line numbers.

Turn on web search for questions about anything after its training cutoff, and ask it to quote sources.

Define your tools explicitly and request JSON-schema structured output when chaining Kimi K3.1 Pro into an agent loop.

Attach multiple images (screenshots, diagrams, charts) in one request and reference them by position in the prompt.

Raise the reasoning effort for hard coding or math tasks; lower it for quick extraction to save tokens and latency.

Reuse a stable system prompt so Venice's cached-input rate ($0.25/1M) kicks in instead of full-price input tokens.

Alternatives

How it compares

ModelBest forContextOpen weightsPrice (Venice)
Kimi K3.1 ProAgentic coding & long-context research1M tokensNo$2 in · $10 out / 1M
Qwen 3.8 MaxBalanced frontier chat & reasoning1M tokensNo$2.50 in · $7.50 out / 1M
Grok 4.7Reasoning with a real-time bent500K tokensNo$2.27 in · $6.80 out / 1M
Gemini 3.8 FlashHigh-volume, cost-sensitive workloads1M tokensNo$0.94 in · $4.69 out / 1M
Muse WonderPeak proprietary quality1M tokensNo$5 in · $25 out / 1M
Venice Uncensored 1.2Uncensored, low-cost chat128K tokensYes$0.20 in · $0.90 out / 1M

Kimi K3.1 Pro is the right pick when you need frontier agentic coding and million-token analysis under open weights and zero-retention privacy. If cost dominates, Gemini 3.8 Flash; if uncensored output dominates, Venice Uncensored 1.2.

Pricing

What it costs on Venice

Billed per token on Venice: $2.50 per 1M input tokens and $12.50 per 1M output tokens.

Input / 1M tokens
$2.50
Per 1M tokens
Output / 1M tokens
$12.50
Per 1M tokens
Cached input / 1M
$0.25
Per 1M tokens

New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.

Getting a key

From nothing to a first call

  1. 01

    Create a key in API settings. Nothing else is required to start.

  2. 02

    Export it as VENICE_API_KEY so the snippets above run unedited.

  3. 03

    Point an existing OpenAI client at https://api.venice.ai/api/v1. The scheme is part of the value: an OpenAI client given a bare host does not resolve it.

  4. 04

    Pass e2ee-kimi-k3-1-pro-p as the model and send the request.

FAQ

Frequently asked questions

Kimi K3.1 Pro is Moonshot AI's flagship open-weight model: a 2.8T-parameter mixture-of-experts (104B active per token) with a 1M-token context window, native image understanding, tool calling, and web search. It was post-trained from Kimi K3 with large-scale reinforcement learning and targets agentic coding, financial research, and long-document reasoning.

Venice bills Kimi K3.1 Pro per token: $2.50 per 1M input tokens, $12.50 per 1M output tokens, and $0.25 per 1M for cached input. There is no subscription required — you pay only for what you use.

Kimi K3.1 Pro has open weights, published on Hugging Face under the Kimi K3.1 Pro License — a permissive licence that allows use, modification, and redistribution, with a separate-agreement clause only for Model-as-a-Service providers above $20M in annual revenue. Using it on Venice is pay-per-token, not free.

Yes. On Venice, Kimi K3.1 Pro supports tool use / function calling, structured JSON-schema output, web search, image input including multiple images per request, and extended reasoning. Video input is not offered through the Venice endpoint.

Venice runs Kimi K3.1 Pro under its private tier with zero retention: prompts and outputs are not stored, profiled, or used for training. Inference is end-to-end encrypted and executes inside a trusted execution environment, so even Venice cannot read your session content.

Both offer 1M-token contexts at the same $2.50/1M input price on Venice. Kimi K3.1 Pro is the stronger agentic coder and the only one with open weights; Qwen 3.8 Max costs less per output token ($7.50 vs $12.50). Choose Kimi for code and agent pipelines, Qwen for high-volume generation on a budget.

Kimi K3.1 Pro supports a 1,000K (roughly one million) token context window with up to 128K output tokens per response on Venice. Its architecture uses Gated DeepSeek Sparse Attention alongside KDA layers to keep long-context inference efficient.

Moonshot AI announced Kimi K3.1 Pro on October 23, 2026, and published the full model weights on November 4, 2026. It has been available on Venice since November 2026.

Yes. It uses Moonshot's MoonViT-V2 vision encoder (401M parameters) for native image understanding, and the Venice endpoint accepts multiple image inputs per request — useful for comparing charts, screenshots, or document pages in a single prompt.

Run Kimi K3.1 Pro privately

One key, free to start, no credit card.

Start chat