Short answer
Kimi K3.1 Pro's weights are on Hugging Face at moonshotai/Kimi-K3.1-Pro. According to the model card, it's a Mixture-of-Experts model with 2.8T total and 104B activated parameters, a 1,048,576-token context, text and image input, and weights trained with MXFP4 quantization. The checkpoint is about 1.56 TB.
- License: the Kimi K3.1 Pro License is permissive — use, modify, fine-tune and redistribute — with two conditions that only hit very large businesses (details below).
- Hardware: vLLM's recipe runs it on one 8-GPU tensor-parallel group: two GB300 trays, one B300 node or one AMD MI355X node. This is data-center hardware, not a workstation.
- If you just want to use it: call
kimi-k3.1-proover the API. Moyi API supports it with the same key you use for Claude and GPT, in Claude Code, Codex or any OpenAI-compatible client. Moonshot's list price is $2 per million input tokens and $10 per million output tokens.
New accounts start with a $0 balance — top up before your first request. Top up from $5. If a request returns 402 (400 on the OpenAI format) with "wallet balance is insufficient", your balance is empty — your base URL and key are fine. Top up and retry.
Use Kimi K3.1 Pro over the API — get a key
What Moonshot released
From the model card:
| Kimi K3.1 Pro | |
|---|---|
| Architecture | Mixture-of-Experts, 896 routed experts, 16 selected per token, 2 shared |
| Parameters | 2.8T total, 104B activated |
| Context length | 1,048,576 tokens |
| Attention | Kimi Delta Attention plus Gated DeepSeek Sparse Attention (replacing K3's full-attention layers for long context) |
| Modality | Text and image (MoonViT-V2 vision encoder, 401M parameters) |
| Weights | MXFP4 weights / MXFP8 activations, from quantization-aware training; about 1.56 TB on Hugging Face |
| Recommended engines | vLLM, SGLang, TokenSpeed |
Moonshot describes it as its most capable 3T-class model, post-trained from Kimi K3 with large-scale reinforcement learning, with the biggest gains in coding, 3D creation, financial research and factual accuracy (tech blog). Reasoning is always on; depth is set with reasoning_effort (low, high, max, default max).
The license in plain words
The Kimi K3.1 Pro License grants free use, copying, modification, fine-tuning and redistribution, as long as you keep the notice. Two conditions apply only at scale:
- Large "model as a service" businesses: if you sell API-style access to model inference or fine-tuning and your group's revenue exceeds $20 million over any 12 months, you need a separate agreement with Moonshot before commercial use.
- Very large products: if a commercial product built on it has more than 100 million monthly active users or more than $20 million in monthly revenue, it must display "Kimi K3.1-Pro" prominently in its interface.
Neither applies to internal use that doesn't expose the model to third parties, or to use through Moonshot's official products and certified inference partners. Read the license itself before you ship; this is a summary, not legal advice.
What it takes to run it yourself
- Disk and memory: the checkpoint alone is about 1.56 TB, and you need room for the KV cache on top for long contexts.
- GPUs: vLLM's baseline is one tensor-parallel group of 8 GPUs — two GB300 trays or one B300 node — and one MI355X node on AMD. The recipe notes that throughput wasn't tuned.
- Software: a recent vLLM (the recipe lists 0.34.0) with
--trust-remote-codeand thekimi_k3tool-call and reasoning parsers; SGLang and TokenSpeed have their own guides. - Multi-turn detail: the model card says the full assistant message, including
reasoning_contentandtool_calls, must be passed back unchanged in later turns.
Self-hosting vs the API
| Run the weights yourself | Call the API | |
|---|---|---|
| Up-front cost | An 8-GPU Blackwell-class or MI355X node | None |
| Cost model | Hardware and power, whether busy or idle | Per token: $2 in / $10 out per 1M at Moonshot's list price |
| Setup and upkeep | Serving stack, upgrades, monitoring | A base URL and a key |
| Scaling | Buy or rent more nodes | Handled by the provider |
| Data control | Stays on your machines | Goes to the API provider |
| Fine-tuning | Possible | Not on Moyi API |
Self-hosting makes sense when you have strict data-residency needs, steady high volume, or want to fine-tune. For trying the model, coding with it, or variable traffic, the API is cheaper and faster to start.
Use Kimi K3.1 Pro over the API
Claude Code. Point it at Moyi API and set the model:
export ANTHROPIC_BASE_URL=https://api.moyi-api.com
export ANTHROPIC_AUTH_TOKEN="your Moyi API key"
export ANTHROPIC_MODEL=kimi-k3.1-pro
Type the ID exactly — kimi-k3.1-pro, without a [1m] suffix, which currently returns an error. More detail: Kimi K3.1 Pro in Claude Code.
OpenAI SDK. Change only base_url and the model:
from openai import OpenAI
client = OpenAI(base_url="https://api.moyi-api.com/v1", api_key="sk-...")
response = client.chat.completions.create(
model="kimi-k3.1-pro",
messages=[{"role": "user", "content": "Write a Python function that parses ISO 8601 durations."}],
)
print(response.choices[0].message.content)
The model page with Moyi API's current price is at /models/kimi-k3.1-pro. You top up from $5 by card, with no subscription.
FAQ
Is Kimi K3.1 Pro open source?
Its weights are open on Hugging Face under the Kimi K3.1 Pro License, which allows commercial use, modification and redistribution with two conditions for very large businesses.
How big is Kimi K3.1 Pro?
2.8T total parameters with 104B activated per token; the MXFP4 checkpoint is about 1.56 TB.
Can I run Kimi K3.1 Pro on a single GPU or a Mac?
Not the full model. vLLM's recipe uses eight data-center GPUs (B300, GB300 or MI355X class) in one tensor-parallel group.
Is the API version the same model?
Yes: kimi-k3.1-pro on Moonshot's API and on Moyi API is Kimi K3.1 Pro.