November 8, 2026 · 5 min read
Kimi K3.1 Pro and MiniMax M3.1 Flash are live on Litepod
Two new models are available on the Litepod API today: Kimi K3.1 Pro from Moonshot AI and MiniMax M3.1 Flash from MiniMax. Both take text and images, call tools, and read up to about a million tokens at a time. They sit at opposite ends of the price range, so most teams will want one of each: the heavyweight for work that has to be right, the flash model for work that has to be cheap.
At a glance
| Kimi K3.1 Pro | MiniMax M3.1 Flash | |
|---|---|---|
| Model ID | moonshotai/kimi-k3.1-pro | minimax/minimax-m3.1-flash |
| Size | 2.8T total, 104B active per token | 428B total, about 23B active per token |
| Context window | 1,048,576 tokens | 1,048,576 tokens |
| Capabilities | Function calling, vision, fine-tuning | Function calling, vision, reasoning |
| Input | $2.00 per 1M tokens | $0.15 per 1M tokens |
| Cached input | $0.20 per 1M tokens | $0.015 per 1M tokens |
| Output | $10.00 per 1M tokens | $0.60 per 1M tokens |
Cached input costs 90% less on both. If every request starts with the same system prompt, tool list or document, you pay a tenth of the input price for that repeated part.
Kimi K3.1 Pro: for work that has to be right
Kimi K3.1 Pro is a mixture-of-experts model: 2.8 trillion parameters in total, with 104 billion used for each token. You get the quality of a very large model without paying for every parameter on every request. It is built for agents that use tools, long coding sessions, and tasks that mix documents and images.
| Benchmark | Kimi K3.1 Pro | Compared with | Source |
|---|---|---|---|
| Intelligence Index | 55.1 | 43.6 Kimi K3 | Artificial Analysis |
| Terminal-Bench 4.0 | 67.2% | 59.6% Claude Opus 5.5 | Artificial Analysis |
| SWE Marathon | 51.7 | 46.3 Claude Opus 5.5 | Moonshot AI |
| Humanity's Last Exam, with tools | 70.0 | 67.7 Claude Opus 5.5 | Moonshot AI |
| OfficeQA Pro v2 | 64.4 | 62.2 Claude Opus 5.5 | Moonshot AI |
The jump over the previous Kimi K3 is large: 55.1 against 43.6 on the Artificial Analysis index, and 67.2% against 12.6% on Terminal-Bench 4.0, a test of working in a real command line.
MiniMax M3.1 Flash: for work that has to be cheap
MiniMax M3.1 Flash activates about 23 billion of its 428 billion parameters per token. That keeps replies quick and the bill small, which suits high-volume jobs: classification, extraction, summaries, routing, and chat. At $0.15 per million input tokens it is among the lowest-priced models with a million-token window on the platform.
| Benchmark | MiniMax M3.1 Flash | Compared with | Source |
|---|---|---|---|
| Intelligence Index | 43 | 29 MiniMax M3, 41 Gemini 3.8 Flash | Artificial Analysis |
| Cost to run the index, per task | $0.22 | $1.24 Gemini 3.8 Flash | Artificial Analysis |
| Humanity's Last Exam | 48.0% | 39.0% MiniMax M3 | Artificial Analysis |
| Output speed, first token | 93 tokens/s, 1.38 s | n/a | Artificial Analysis |
| Automation-Bench | 62.8 | 50.3 Claude Opus 5 | MiniMax |
On the same independent index it scores 43, up from 29 for the earlier MiniMax M3 and ahead of Gemini 3.8 Flash at 41, while costing about a fifth as much to run per task.
Know the trade-offs
Benchmarks are a starting point, and a few of these numbers come with fine print:
- Who ran the test matters. Rows marked Artificial Analysis are independent runs. Rows marked Moonshot AI or MiniMax are published by the model makers on their own setups. The two can disagree: on Terminal-Bench 4.0, Moonshot reports 65.7 against Claude Opus 5.5's 66.4, while Artificial Analysis measures 67.2% against 59.6%.
- Kimi K3.1 Pro is not the fastest. Artificial Analysis measured about 57 output tokens per second, which is slow for interactive chat. It is a better fit for jobs where quality counts more than speed.
- It trails the top closed model overall. Its Intelligence Index of 55.1 is behind Claude Opus 5.5 at 57.6.
- MiniMax M3.1 Flash is not the top flash model. DeepSeek V4.1 Flash scores 45 on the same index and generates faster, and it is also on Litepod. M3.1 Flash is stronger on knowledge and agent tasks.
- Check it on your own data. On a factual-recall test, M3.1 Flash gave a wrong answer instead of declining 92.8% of the time it did not know. If your app depends on accurate facts, pair it with retrieval or check its output.
Which one should you pick?
- Choose Kimi K3.1 Pro for coding agents, multi-step tool use, and long documents or codebases where a wrong answer costs more than the tokens.
- Choose MiniMax M3.1 Flash for volume: millions of short requests, background pipelines, and anything where price per call is the deciding factor.
- Use both: route easy requests to the flash model and send the hard ones to Kimi. It is a one-line change because both sit behind the same endpoint.
Try it
There is nothing new to install. Litepod is OpenAI-compatible, so point your existing client at https://api.litepod.io/v1 and set the model field.
curl https://api.litepod.io/v1/chat/completions \
-H "Authorization: Bearer $LITEPOD_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "minimax/minimax-m3.1-flash",
"messages": [{"role": "user", "content": "Summarise this ticket in one line."}]
}'Create a key from the dashboard, compare every model on the models page, or read the quickstart. To try the platform first, Nemotron Lightning 3.5 is free: 50 requests a day.
Benchmark figures are from Artificial Analysis and from the Moonshot AI and MiniMax launch announcements, as of November 2026, and may change. Prices are in US dollars per million tokens and always match the models page.