Kimi K3 and Kimi K3.1 Pro: benchmarks, features and access on QDivZero
Published 10 November 2026 by QDivZero Team10 min read
Discover the models and usage included in the plan
Explore Unlimited Basic and ProKimi K3 and Kimi K3.1 Pro are now available on QDivZero, through both the Unlimited Basic subscription and pay-as-you-go Public Models. Connect them to your applications through the same OpenAI-compatible API, choosing whether a request uses the plan’s AI Credits or your euro balance.
These Moonshot AI models focus on coding, reasoning and agentic work. We explore their main features, published benchmark results and how to access both models on QDivZero.
Unlimited Basic stays at €14.95/month, VAT included, with two unlimited Qwen models subject to fair use and 5,000 monthly AI Credits shared across eleven premium models, including both Kimi versions. You can also use Kimi through balance without a subscription. See the current access options and terms.
In this article
- What Kimi K3 and Kimi K3.1 Pro are.
- Features and use cases.
- Kimi K3 and Kimi K3.1 Pro benchmarks.
- How they work in Unlimited Basic.
- Public Model pricing.
- Calling the Kimi API on QDivZero.
- Choosing a model and access option.
- Frequently asked questions.
What are Kimi K3 and Kimi K3.1 Pro?
Kimi K3 is a multimodal open-weight model from Moonshot AI. Its technical report describes a Mixture-of-Experts architecture with 2.8 trillion total parameters and 104 billion active parameters, native vision and a one-million-token context window. Activating a subset of parameters distributes work across specialised experts.
Kimi K3.1 Pro extends that family through larger-scale reinforcement-learning post-training. Its official model card retains 2.8 trillion parameters, 104 billion active parameters and a 1,048,576-token context, while replacing K3’s Gated MLA layers with Gated DeepSeek Sparse Attention.
This article uses Kimi K3.1 Pro, the official name, to distinguish it from informal names such as “Kimi K3 Pro”. The two versions have different identifiers on QDivZero.
The published context length describes the underlying model. When preparing an integration, check the limits and input formats supported by the specific endpoint in the platform. A larger context can also increase usage: text resent with each turn counts as input.
Kimi features: coding, agents and document analysis
Coding and multi-step tasks
The official Kimi K3 announcement highlights long engineering sessions, repository navigation and terminal-tool orchestration. These capabilities suggest practical experiments such as locating a bug, proposing an implementation and checking it with tests.
A useful first trial is to give both models the same problem, relevant files and an acceptance criterion. Measure whether the solution works and how many iterations it takes. That evidence will help you choose more effectively than a convincing-looking response alone.
Vision and working with extensive information
The Kimi K3 model card describes text and image understanding alongside video capabilities. This combination opens up experiments such as reviewing an interface screenshot or reasoning over a chart. For documents, preserve page references, units and citations when preparing inputs so that you can check the answer afterwards.
Input formats depend on the service you use. For an API integration, prepare data according to the endpoint’s supported schema and use the inference documentation as a reference.
What Kimi K3.1 Pro adds
The Kimi K3.1 Pro announcement focuses on coding, 3D creation and financial research. Its examples include code for CAD tools and tasks combining document retrieval, evidence checks and numerical reasoning.
These capabilities need a suitable environment. A model can propose a query, script or tool call; your application or agent provides permissions and executes those actions. When designing a workflow, define what it can read, what it can change and how you will review the result. You can explore this kind of work with QDivZero Harness, our application for working with models and projects.
Kimi K3 and Kimi K3.1 Pro benchmarks: results and methodology
The following figures come from the linked publications. Moonshot’s results are developer-published evaluations, rather than measurements performed by QDivZero. Each benchmark measures a particular task and may use different tools, budgets and agent environments.
Kimi K3: reasoning, coding and vision
Source: official Kimi K3 evaluation table. Moonshot reports max reasoning effort and temperature 1.0; Terminal-Bench uses the Kimi Code environment. Figures follow each benchmark’s scoring scale, and tool-based tests include the execution environment.
Kimi K3.1 Pro compared with other models
Source: Kimi K3.1 Pro evaluation table and notes. K3.1 Pro’s published configuration is max effort, temperature 1.0 and top-p 0.95 unless otherwise specified. The notes identify Claude Code as the environment for Terminal-Bench 4.0 and SWE Marathon v1.1; peer results use the sources and configurations listed by Moonshot.
Benchmark versions matter: Terminal-Bench 2.1 and Terminal-Bench 4.0 are different evaluations. K3’s 88.3 and K3.1 Pro’s 65.7 should therefore be read within their respective tables.
An external comparison on data analysis
Plotly tested both Kimi models with its data analytics exam: K3 passed 35 of 43 questions, while K3.1 Pro passed 42 of 43. In that evaluation, Pro found all five hidden patterns and all three data quality issues planted by the authors.
Plotly also observed longer time per question and greater output usage with Pro. Its costs and latencies describe its run through another provider. To budget a QDivZero integration, use our rates and measure usage on your own task.
The practical approach is to evaluate quality, cost per completed task and response time together. An automation that completes more steps can consume more tokens, even when its per-token rate is lower.
Kimi K3 and Kimi K3.1 Pro in Unlimited Basic
The Unlimited Basic subscription gives you access to both Kimi models through 5,000 monthly AI Credits shared across the eleven premium models. Allocate them to Kimi, DeepSeek, GLM, embeddings, transcription and decisions as your work requires.
Both Kimi models consume AI Credits. Unlimited usage subject to fair use applies to Qwen 3.8 27B and Qwen 3.6 35B-A3B, which remain available when premium credits run out. The subscription stays at €14.95/month, VAT included, with no minimum commitment.
How many AI Credits do they use?
Plan rates verified on 10 November 2026. See the Unlimited catalog and usage estimates. Input and output costs are added to calculate each request’s usage.
For example, 90,000 input tokens and 10,000 output tokens use 251.16 AI Credits with K3 or 180.32 with K3.1 Pro. Under these rates, Pro costs less per token; the final task cost depends on how many tokens it needs.
What can you do with 5,000 AI Credits?
These QDivZero estimates allocate the entire allowance to one model, with 90% input tokens and 10% output tokens:
These are token budgets for each scenario, rather than fixed sizes for all questions or projects. Resent context counts as input. Rows and columns are not additive: every estimate uses all 5,000 AI Credits, and the allowance is shared with the other premium models.
Credits renew with the monthly cycle and unused credits expire. When they run out, you can wait for renewal or explicitly select a public model to use balance. The Unlimited documentation explains plan management and limits.
Kimi K3 and Kimi K3.1 Pro Public Model pricing
If you prefer usage-based billing, both are available as Public Models. Top up your balance and requests deduct euros according to input and output tokens. You can use this option without Unlimited Basic or combine both in the same account.
Source: QDivZero’s public rate catalog, checked on 10 November 2026. Figures are rounded to four decimal places; calculations use the full rates. Check current availability and rates on the pricing page.
For the example with 90,000 input tokens and 10,000 output tokens, the approximate cost is €0.2791 with Kimi K3 and €0.2004 with Kimi K3.1 Pro.
Euro balance and AI Credits are separate allowances. The selected identifier determines which one you use. A request to the public variant uses balance even when you also have an active subscription.
How to use the Kimi K3 API on QDivZero
These access options provide prepared endpoints: you can start without downloading weights or deploying a GPU. Create an API key, prepare your subscription or balance and use the base URL https://api.qdiv0.com/v1.
Subscription and public model identifiers
Switch versions while keeping the request structure. In OpenAI-compatible clients, configure the base URL, your QDivZero key and one of these complete identifiers. The quickstart documentation walks through the connection.
cURL example with Unlimited Basic
Set QDIV0_API_KEY in your terminal and send a Chat Completions request. This example selects Kimi K3.1 Pro with AI Credits:
1curl https://api.qdiv0.com/v1/chat/completions \
2 -H "Authorization: Bearer $QDIV0_API_KEY" \
3 -H "Content-Type: application/json" \
4 -d '{
5 "model": "subscription:moonshotai/kimi-k3.1-pro",
6 "messages": [
7 {
8 "role": "user",
9 "content": "Propose a plan to diagnose a slow API. Explain what to measure, how to find the bottleneck and how to validate the improvement."
10 }
11 ],
12 "max_tokens": 4096,
13 "stream": false
14 }'
15To use balance, change model to public:kimi-k3.1-pro. To compare K3, use its identifier from the table. Review recorded usage and adjust the output budget and context for your task.
An agent that browses, runs code or queries a database also needs those tools in your application. The API provides model access; your integration defines the workflow. Check the API reference before adding specific parameters or input formats.
Kimi K3 vs Kimi K3.1 Pro: choosing for your application
Kimi K3.1 Pro is a useful starting point for a new evaluation if you prioritise its published coding and data analysis results. At the time this article was published, it also has lower per-token rates in both QDivZero access options. Measure total usage and latency on a representative task.
Kimi K3 remains useful as a comparison or for an already validated workflow. Keep a small set of real cases—a code bug, a document question and an analysis task—and repeat them with both models.
To choose an access option, start with expected usage: Unlimited Basic brings several models and a monthly allowance into a fixed subscription; Public Models lets you manage spending through balance and usage-based billing. You can use both and select the appropriate identifier for each request.
Explore Unlimited Basic models, check public rates or open QDivZero to prepare your first trial.
Frequently asked questions about Kimi K3 and Kimi K3.1 Pro
Are Kimi K3 and Kimi K3.1 Pro included in Unlimited Basic?
Yes. Both use the monthly 5,000 AI Credits shared across the eleven premium models. The plan costs €14.95/month including VAT and also includes two unlimited Qwen models subject to fair use.
Does the subscription include unlimited Kimi tokens?
Kimi uses AI Credits according to input and output tokens. Unlimited Basic’s models with unlimited usage subject to fair use are Qwen 3.8 27B and Qwen 3.6 35B-A3B.
Can I use the Kimi API without a subscription?
Yes. Select public:kimi-k3 or public:kimi-k3.1-pro and use prepaid balance. The subscription: identifiers use the plan’s AI Credits.
What is the difference between Kimi K3 and Kimi K3.1 Pro?
K3.1 Pro advances the K3 family through post-training. When choosing on QDivZero, compare published evaluations and test both with your data, measuring quality, tokens and time per task.
Do the benchmarks show QDivZero’s speed?
The tables above report task results under their authors’ evaluation conditions. To measure speed and usage on QDivZero, run your own workload against our endpoints and review the results.
Is Kimi open-source?
Moonshot publishes model weights under its own licence. We use open-weight to describe this distribution. If you plan to download, modify or redistribute the models, check the terms in the official K3 and K3.1 Pro model cards.
Sources and verification date
The information in this article was checked on 10 November 2026. Benchmarks refer to the cited versions and conditions; prices and availability may change.
