Kimi K3.1 Pro on Felo API: Capabilities, Benchmarks, Pricing, and Setup
A detailed guide to Kimi K3.1 Pro on Felo API: its 1M-token context, visual input, coding and finance results, pricing, and API setup.
Kimi K3.1 Pro is now available through Felo API. If you are considering it for a coding agent, a visual workflow, or research across a large document set, this guide covers the details you need before writing your first integration: what changed from K3, what the published evaluations actually show, which inputs Felo supports, and how to choose between its two provider routes.
The public Felo model ID is kimi-k3.1-pro. You can inspect it in the Felo model catalog and test it in the Playground.
Kimi K3.1 Pro at a glance
| Detail | What is available now |
|---|---|
| Model | Moonshot AI's first K3.1-series flagship, a post-trained version of Kimi K3 |
| Felo model ID | kimi-k3.1-pro |
| Context window | 1,048,576 tokens (about 1 million) |
| Input | Text, images, and video |
| Output | Text; it can write code for a 3D workflow, but the model call does not directly return a 3D asset or generated video |
| Reasoning | Enabled by default; Felo lists max, high, low, and none effort levels, with max as the default |
| Recommended Felo protocol | Messages, at POST https://openapi.felo.ai/api/v1/messages |
| Provider routes on Felo | Moonshot AI MXFP4 and Moonshot AI Highspeed |
These are the fields in Felo's current model listing. A large context window is a capacity limit, not a promise that every fact in a million-token prompt will be recalled perfectly. For production work, test the length, input format, and reasoning setting you actually plan to use.
What changed from Kimi K3?
Moonshot describes K3.1 Pro as a post-trained version of K3, with substantially more post-training compute and reinforcement learning across many downstream tasks. That distinction matters. The public claim is about improved behavior on coding, agentic, visual, and knowledge-work tasks; it is not a claim that K3.1 has a newly published base architecture. Felo's K3.1 Pro model page summarizes the model and its supported API route.
K3 itself is a 2.8-trillion-parameter, 3T-class mixture-of-experts model. Its published design uses Kimi Delta Attention and Attention Residuals, with native visual understanding and a one-million-token context window. Because K3.1 Pro builds on K3, this is useful architectural background; Moonshot has not yet released the full K3.1 weights or a separate architectural specification. Explore K3's architecture and specifications on Felo.
The practical change is in what the model can carry through to completion. Moonshot reports gains in long-horizon coding, 3D modeling, financial research, and factual accuracy on its stress test. K3.1 Pro also has a Highspeed serving option that Moonshot says delivers roughly 2–3 times the speed of the regular version with the same model capability. The faster route has a higher token price, covered below.
The published entry price has changed too. Felo lists K3 at $3 per million input tokens and $15 per million output tokens. K3.1 Pro's MXFP4 route is listed at $2 and $10 respectively. That makes the newer model's lower-priced route one-third less expensive per uncached input or output token at today's listed rates. Compare the exact channels and latency you intend to use before treating that as a full workload saving.
Where the model is strongest
Long-horizon coding and agents
K3.1 Pro is meant to work beyond a single code snippet. Moonshot's examples include multi-file development, debugging, front-end builds, game-engine work, and tool-assisted tasks that require repeated inspection and revision. Its own Kimi Code Bench 3.0 Long-horizon result rose from 55.95% for K3 to 61.2% for K3.1 Pro. On Terminal-Bench 4.0, K3.1 Pro scored 65.7% in Moonshot's published evaluation. For an engineering team, the useful question is whether it can change a repository without losing the thread: read the right files, propose a patch, run the relevant checks, and recover when a test fails. Give it a representative issue and grade the final diff and test results. A polished plan by itself is not evidence that the task is complete.
Images, video, and 3D work
The Felo route accepts text, image, and video input and returns text. That makes K3.1 Pro relevant to tasks such as explaining a screenshot, comparing frames, reasoning over a visual reference, or writing code and scripts for a 3D modeling pipeline. It does not turn the model endpoint into an image, video, or 3D-file generator. To produce an asset, connect the model's output to your own tools or an appropriate generation workflow.
Moonshot reports 92.7% on BenchCAD with tools, a test of reconstructing 3D objects from multiple rendered views through generated CAD code. Its published example set also includes Blender and CAD scripting, interactive Three.js scenes, and work in Unity, Unreal Engine, and Godot. Those examples show the kind of work to evaluate; they do not mean every tool is included with a plain API request.
Financial research and modeling
Moonshot reports 64.4 on OfficeQA Pro v2, which tests analysis across financial documents, including retrieval, cross-document checks, and numerical reasoning. On Moonshot's internal evaluations, K3.1 Pro scored 70.4 versus 58.8 for K3 in financial research and 57.6 versus 50.4 in financial modeling. These are task-specific scores, not investment returns or measures of accuracy on every financial question.
This is a good fit to test with a controlled corpus: annual reports, filings, earnings materials, and a question that requires the model to cite exact passages and show calculations. Keep the source documents and arithmetic available for review. In a workflow that makes financial decisions, independent verification is still necessary.
Long documents and multi-step research
The one-million-token window lets an application submit far more material than a typical short chat request: code, transcripts, reports, images, and conversation history can share the same context. The model can also participate in multi-step agent workflows through tool calls. But context capacity and agent behavior are separate things. A long prompt still needs a clear task, a document map, and a way to check the answer; an agent still needs the developer to provide tools and handle their results.
For research, ask the model to identify the relevant source first, then answer with page or section references and a separate list of uncertain points. That structure makes errors easier to catch than a long, fluent summary.
Moonshot and Felo also describe multi-agent orchestration that can scale to hundreds of parallel subagents. That is an agent-system capability built around the model, not something a single API response creates on its own. A developer still has to coordinate workers, tools, budgets, and the final review step. See the Felo model description.
How to read the benchmark numbers
Moonshot's announcement contains many evaluations. The following selection covers distinct types of work without treating unlike tests as one ranking.
| Evaluation | K3.1 Pro | Comparison reported by Moonshot | What it tests |
|---|---|---|---|
| Terminal-Bench 4.0 | 65.7% | Claude Opus 5.5: 66.4%; GPT-6 Astra: 58.2% | Coding agents on terminal tasks with task-specific tests |
| Kimi Code Bench 3.0 Long-horizon | 61.2% | K3: 55.95% | Moonshot's internal multi-step coding benchmark |
| BenchCAD, with tools | 92.7% | Claude Opus 5.5: 96.2%; GPT-6 Astra: 95.9% | CAD code generation from multi-view renderings |
| Financial research, internal | 70.4 | K3: 58.8 | Evidence-grounded financial research |
| Financial modeling, internal | 57.6 | K3: 50.4 | Workbook formulas, financial logic, and final model quality |
| Kimi Hallu Bench Hard | 8.3% claim-level hallucination rate | K3: 23.4%; lower is better | A deliberately difficult factual-accuracy stress test |
All values are from Moonshot AI's K3.1 Pro technical announcement, not Felo-run evaluations. Unless stated otherwise, Moonshot reports K3.1 Pro results at max reasoning effort, temperature 1.0, and top-p 0.95. The harness varies by benchmark, and some comparisons draw on other providers' published results. In particular, Kimi Code Bench and the two finance scores above are internal evaluations. Use them to choose tests for your own workload, not to assume the same pass rate in production. Felo also reproduces a selection of these results on its model page.
The hallucination figure needs its own warning. 8.3% is the claim-level rate on Kimi Hallu Bench Hard, a stress test built to expose mistakes. Moonshot explicitly says it is not the model's everyday error rate. It is evidence of improvement over K3 under that test, not a reason to skip source checking.
Pricing and provider choice on Felo API
Felo lists two runnable Moonshot AI channels for kimi-k3.1-pro. These are the published rates per one million tokens, in US dollars:
| Felo channel | Input | Output | Cache read |
|---|---|---|---|
moonshotai-mxfp4 | $2 | $10 | $0.20 |
moonshotai-highspeed | $4 | $20 | $0.20 |
The live model page is the source for Felo's rates. Highspeed is listed as the default channel in Felo's catalog. If your budget or latency target depends on a specific route, send X-Felo-Channel explicitly rather than assuming which provider a request will use. Felo's provider-routing guide explains the header and its fallback behavior.
For example, a request with 200,000 uncached input tokens and 8,000 output tokens would cost about $0.48 on MXFP4 (0.2 × $2 + 0.008 × $10) or $0.96 on Highspeed (0.2 × $4 + 0.008 × $20). If 100,000 input tokens qualify for the published cache-read rate, that portion costs $0.02 on either channel. A cache hit is not guaranteed; it depends on the upstream service and your request. Felo's pricing guide explains how token and cache charges are calculated.
These are estimates for model tokens. Check current rates, usage records, and any other billable tools before using the numbers in a production budget. Felo API uses prepaid service credits; a subscription is not required for API usage. Read the billing guide.
Start using Kimi K3.1 Pro on Felo
- Create a Felo API key and add service credits if your account needs them.
- Try one prompt in the Playground. Use material similar to what your application will send.
- Move the prompt into your client using the public model ID. The Messages endpoint is the recommended protocol on Felo's model page.
The request below pins the lower-priced MXFP4 channel, so the pricing choice is explicit:
export FELO_API_KEY="YOUR_API_KEY"
curl https://openapi.felo.ai/api/v1/messages \
-H "Authorization: Bearer $FELO_API_KEY" \
-H "Content-Type: application/json" \
-H "X-Felo-Channel: moonshotai-mxfp4" \
-d '{
"model": "kimi-k3.1-pro",
"max_tokens": 1024,
"messages": [
{
"role": "user",
"content": "Review this TypeScript refactor plan. Identify missing tests, risky changes, and the first file you would inspect."
}
]
}'
To select the faster route, change that header to X-Felo-Channel: moonshotai-highspeed. To let Felo choose, omit the header. Felo accepts the same routing header on its Chat Completions, Responses, and Messages model endpoints. An explicitly selected channel does not automatically fail over to another provider; your application must retry with a different channel if that is what you want. See Felo's routing documentation.
SDKs, streaming, tools, and reasoning
Felo provides three compatible model-calling surfaces. Use the one your application already speaks:
| Client style | Endpoint | SDK base URL |
|---|---|---|
| OpenAI Chat Completions | /api/v1/chat/completions | https://openapi.felo.ai/api/v1 |
| OpenAI Responses | /api/v1/responses | https://openapi.felo.ai/api/v1 |
| Anthropic Messages | /api/v1/messages | https://openapi.felo.ai/api |
For Anthropic-compatible SDKs, the base URL stops at /api; the SDK adds /v1/messages. The Messages protocol requires max_tokens. Add "stream": true to receive server-sent events, and wait for the protocol's completion event rather than treating a dropped connection as a finished answer. Felo's protocol guide covers the request shapes and streaming behavior.
These are platform-level protocol options. Confirm that the parameter and modality combination you need works on the K3.1 Pro route before migrating an existing integration.
Felo lists tools, tool_choice, structured-output parameters, and reasoning controls for this model. The supported reasoning-effort values are max, high, low, and none, with max as the default. Model and protocol support can differ, so test the exact parameter combination your client needs before relying on it. The published benchmark results largely use max; lowering effort is a quality-and-cost experiment, not a way to reproduce those scores. Check the current model capabilities.
If your workflow needs fresh web data, a calculator, a repository shell, or a CAD program, provide and operate those tools in your application. Calling kimi-k3.1-pro alone does not execute external tools or verify cited claims.
A practical evaluation plan
K3.1 Pro is broad enough that a generic "Which model is best?" test will tell you little. Three small trials produce more useful evidence:
- Coding: Give it a real issue and the relevant repository context. Require a patch, tests, and a short explanation of any remaining risk. Record whether the tests pass and how many tool turns it takes.
- Visual or 3D: Supply a reference image or multiple views and ask for a modeling plan or script. Have a person or renderer check dimensions and geometry rather than judging only the prose.
- Research: Provide a fixed document set and a question with a known answer. Require source locations, calculations, and an explicit "not found" when evidence is missing.
Run each trial on both Felo channels if latency matters. Record input, output, and cache tokens alongside total cost and elapsed time. Keep the same prompts and tools when comparing against K3 or another model. That gives you a decision based on completed work per dollar and minute, not a headline benchmark.
Limits and release status to keep in mind
- Open-weight timing: Moonshot calls K3.1 Pro an open-weight model but says the full weights will be released by November 4, 2026. API access is available now; do not describe the weights as already downloadable. Felo's model page notes the planned release date.
- Output format: Felo currently lists text output, even when the input contains images or video. Asset creation requires a connected tool or another generation endpoint. Felo model listing.
- Benchmark transfer: Internal scores, tool-equipped tests, and long-horizon demos will not map directly to a plain one-shot API call. Match the evaluation setup to your application.
- Accuracy: The improved hallucination result is still an error rate on a stress test. Verify quotations, calculations, and factual claims in high-stakes workflows.
- Context and cost: A one-million-token window can make very large requests possible and expensive. Budget for the actual prompt, generated answer, retries, and repeated context; use cache-read pricing only where usage reports a hit.
Moonshot also describes targeted training controls for cyber-offense capability in the K3.1 Pro release. Teams evaluating defensive security work should test the specific authorized tasks they need rather than assume performance from general coding results.
Frequently asked questions
Is Kimi K3.1 Pro the same as Kimi K3?
No. Moonshot describes K3.1 Pro as a post-trained upgrade of K3 and reports improvements in coding, 3D creation, financial research, and its factual-accuracy stress test. It retains the one-million-token context class and native visual understanding.
Which model ID do I send to Felo API?
Use kimi-k3.1-pro. On Felo, MXFP4 and Highspeed are provider channels for that public model ID; select one with X-Felo-Channel if you need a specific route.
Can it make 3D models or videos directly through the model endpoint?
It can understand visual input and write plans, code, or scripts for 3D and video workflows. Felo currently lists text as this model endpoint's output modality. You need a renderer, design tool, or generation service to produce the asset itself.
Should I use MXFP4 or Highspeed?
Start with MXFP4 when token cost is the main constraint. Test Highspeed when waiting time affects your product or agent loop. Moonshot says Highspeed keeps the same model capability at roughly 2–3× the speed, while Felo's published input and output rates are twice MXFP4's.
Is the full model already available to self-host?
Moonshot says the full K3.1 Pro weights are scheduled for release by November 4, 2026. Check its official release page for the latest status. Felo API access is already available.
Try it with your own workload
Kimi K3.1 Pro is an interesting candidate when the task combines long context, tool use, and visual or document evidence. Its reported gains are substantial, but the quickest way to decide whether they matter to you is to run a representative job, inspect the result, and compare the total cost of the MXFP4 and Highspeed routes.
Explore Kimi K3.1 Pro on Felo API →
This post is also available in 简体中文, 日本語, 한국어, 繁體中文, हिन्दी, Français, العربية, Русский, اردو, Bahasa Indonesia, Deutsch, Tiếng Việt, Türkçe, Italiano, ไทย, Español, বাংলা and Português.