API Market / AI News / Kimi K3.1 Pro goes open-weight: what API users actually get

Kimi K3.1 Pro goes open-weight: what API users actually get

Published: November 5, 2026 · Official

On November 5, 2026, Moonshot AI published the full weights of Kimi K3.1 Pro: 2.8 trillion parameters, 1.56 TB, commercial use permitted with conditions. The model was already reachable through Moonshot's API; now it can also be run in your own infrastructure. Here is what is inside, how it compares, and what is practical for people working through an API.

What was released

The release landed on November 5, 2026: Moonshot AI published the Kimi K3.1 Pro weights on Hugging Face. The download is 1.56 TB, the parameter count is 2.8 trillion. Downloading is free and commercial use is allowed, but with conditions; the exact wording is not disclosed in the available coverage.

Before that, the model existed only as a service: through Kimi.com, the mobile apps, the Kimi Work desktop agent, the Kimi Code tool and Moonshot's API. Open weights add a third option — deployment on your own hardware.

Architecture and capabilities

Kimi K3.1 Pro is a mixture-of-experts model: 2.8 trillion total parameters with roughly 104 billion active per inference. The context window holds up to 1 million tokens, though in Kimi Code that size is available only on higher-tier plans. The model is natively multimodal: it takes text, images and video, and returns text.

According to Moonshot, its strengths are programming, 3D work, financial research, factual accuracy and complex agent tasks. It can run software projects on its own, build games, drive Blender and CAD through its own scripts, and create 3D objects from text descriptions, sketches or reference images. In 3D Make the result can be edited in a browser and downloaded for 3D printing.

The technical report attributes most of the gain to more extensive post-training with reinforcement learning across hundreds of task areas. Moonshot also removed certain training data and tasks to limit offensive cyber capabilities.

How it scores in independent tests

On the Agent Arena leaderboard Kimi K3.1 Pro ranks first among open-weight models and 13th overall, based on more than 17,500 real long-duration agent sessions. The Max variant posted a net improvement of +5.1% against +0.9% for Kimi K3 Max. Bash Reliability reached +7.9%, first among open models; the praise-to-complaint ratio improved to +7.7% from +1.3%; controllability hit +6.5%, also first among open models; task completion confirmation stood at +3.9%, fourth.

Artificial Analysis scored it 55 on the Intelligence Index, fourth place: Claude Opus 5.5 leads at 58, with Claude Sonnet 5.5 and Muse Wonder Full at 56 each. The previous Kimi K3 scored 44. One important detail: on October 27 Artificial Analysis changed its scoring method, and under the new one K3.1 Pro sits at 46 against 58 for Opus 5.5 — a 12-point gap, not 3.

On Terminal-Bench 4.0 the success rate rose from 13% to 67%, the best result in Artificial Analysis testing. On AutomationBench-AA the share of task goals completed within the rules went from 58% to 73%. On GDPval-AA the model scored 1,846 Elo, third behind Claude Opus 5.5 (1,866) and Muse Wonder Full (1,857), a 312-point gain over K3. Humanity's Last Exam moved only from 47% to 52%, while the hallucination rate on AA-Omniscience fell from 53% to 44%.

Pricing, speed and what it means for API users

On Moonshot's API the standard version costs $2 per million input tokens and $10 per million output tokens, one-third less than K3. Cache hits cost $0.20 per million. The Highspeed version, which the company says is two to three times faster, costs $4 and $20 respectively.

The standard version runs at about 74 output tokens per second, against 42 for K3 and 97 for Claude Opus 5.5. Highspeed reaches roughly 129 tokens per second. Task cost stays comparable anyway: $2.11 versus $2.00 for K3 and $5.98 for Claude Opus 5.5 at maximum reasoning effort. The reason is that the model generates an average of 55,000 output tokens per Intelligence Index task, 14% more than K3.

For API users, open weights change little on their own: 1.56 TB of a top-tier model needs serious hardware, and few teams can host it in-house. Open weights are not the same as free access — you still pay for compute. Nor is this open source in the strict sense: training data and methods are not disclosed, and commercial use comes with conditions.

Worth keeping in mind: according to Epoch AI estimates, since January 2026 the best open-weight models have trailed the closed frontier by roughly four months on average. Kimi K3.1 Pro is exactly that kind of fast follower, now very close to the leaders.

Kimi K3.1 Pro vs GPT-6.1 Terra and Mistral

A comparison piece at tech-insider.org places Kimi K3.1 Pro alongside GPT-6.1 Terra and Mistral and mentions a 2.4x gap. Which metric that refers to is not clear from the publication, so the figure deserves caution — it is that source's claim, nothing more.

What is better supported: at release, K3.1 Pro is the strongest open-weight model in the Artificial Analysis table, and Moonshot positions it in coding, 3D and financial analysis against closed flagships such as Claude Opus 5.5 and GPT-6 Astra. There are no concrete measurements against GPT-6.1 Terra or Mistral in the available sources, so that comparison remains positioning for now.

What to use right now

Kimi K3.1 Pro is available in the API Market catalog — the same class of model for code, agents and long documents, without deploying a terabyte of weights yourself. One key, payment in rubles, no VPN.

If the task leans toward general reasoning, GPT-6.1 Terra sits next to it in the catalog, and for the hardest code and analysis there is Claude Opus 5.5. Comparing them on your own prompts is easy: access works the same way for every model, so switching takes a minute. New models appear in the API Market catalog as they become available — keep an eye on it.

Models you can use right now

ModelAPI Market price
Kimi K3.1 Pro
kimi-k3.1-pro		input $0.33, output $1.62 per 1M tokens
GPT-6.1 Terra
gpt-6.1-terra		input $0.10, output $0.50 per 1M tokens
Claude Opus 5.5
claude-opus-5-5		input $0.90, output $4.50 per 1M tokens

Try these models on API Market

One API key for GPT, Claude, Gemini, DeepSeek, Grok and more. OpenAI-compatible endpoint, pay as you go, top-ups from about $0.15.

Get an API keyAll models

Connect in a minute

from openai import OpenAI

client = OpenAI(
    base_url="https://api-market1.shop/v1",
    api_key="sk-your-key",
)

reply = client.chat.completions.create(
    model="kimi-k3.1-pro",
    messages=[{"role": "user", "content": "Hi! Tell me about yourself."}],
)
print(reply.choices[0].message.content)

FAQ

Are the Kimi K3.1 Pro weights really free?

Downloading them is free, and commercial use is permitted with conditions. But running 1.56 TB of weights requires your own hardware, so "free" applies to the files, not the compute.

Can I host the model on my own server?

Technically yes — the weights are on Hugging Face. In practice few can: a 2.8-trillion-parameter model demands serious infrastructure, so most teams will keep using an API.

How far behind Claude Opus 5.5 is Kimi K3.1 Pro?

Per Artificial Analysis, 55 points against 58 under the pre-October 27 method. After the method update, K3.1 Pro sits at 46 against 58 for Opus 5.5 — a 12-point gap.

How is K3.1 Pro different from Kimi K3?

Same architecture: mixture of experts at 2.8 trillion parameters, 104 billion active, up to 1 million tokens of context. The gains came from more extensive post-training: 44 to 55 on the Intelligence Index, 13% to 67% on Terminal-Bench 4.0, plus API prices one-third lower.

Where can I try the model without hosting it?

Kimi K3.1 Pro is in the API Market catalog: one key, payment in rubles, no VPN. You can also compare it with GPT-6.1 Terra and Claude Opus 5.5 on your own tasks there.

Sources

More news

Русская версия