September 14: K3 is now listed

NVIDIA’s Kimi K3 model card and live API catalog list Kimi K3. The model card says the model is ready for commercial or non-commercial use, but the hosted API is governed by NVIDIA’s separate API Trial Terms. Under sections 1.2 and 1.4, the trial API and its generated content are not for production; production use requires a separate subscription from NVIDIA or a service provider.

The trial terms also prohibit confidential, controlled, or sensitive input. Personal data is permitted only when the specific API service expressly allows it. NVIDIA says it collects user content and generated content to improve its products and services, including AI models; service disclosures can set additional retention terms. Do not treat this hosted trial as private coding. The model’s commercial-use description does not change these API conditions. Account access, payment, credits, quota, expiry, and rate limits were not tested. See current Kimi access.

Historical August 1 K2.6 setup

The instructions below preserve the K2.6 example and the catalog observation as of August 1. They do not describe the current K3 listing. Do not rename the old model ID or treat its example as a tested K3 integration.

TL;DR: This August 1 example uses Kimi K2.6, moonshotai/kimi-k2.6; Kimi K3 was absent from the catalog in that historical check. The trial limits above also apply to this historical setup. Use Moonshot’s separately governed API or released K3 weights when you need the newest Kimi flagship, after checking their own terms.

August 1 evidence boundary: The trial restriction is validated in NVIDIA’s public terms, rechecked October 3. The terms say NVIDIA may extend credits; they do not establish a universal credit amount or entitlement. Account access, payment requirement, quota, expiry, and rate limits remain unresolved. Do not label this route universally free, production-ready, private, or suitable for autonomous fallback.

1
2
3

Need the current NVIDIA-hosted Kimi listing? → Check Kimi K3 or K2.6 availability in your account
Need the newest Kimi flagship? → Kimi K3 via Moonshot or released weights
Need production access? → Use a service with production rights under its applicable subscription and terms

Kimi K2.6 identity in the historical August setup

The August 1 catalog check found moonshotai/kimi-k2.6 and did not list Kimi K3; that absence is historical. The live NVIDIA API catalog, checked October 3, now includes both moonshotai/kimi-k2.6 and moonshotai/kimi-k3. The Kimi K2.6 NIM reference describes a 1T-parameter mixture-of-experts model with 32B active parameters, a 256K context window, text and image input, and thinking and instant modes.

That identity matters. Kimi K3 is now listed, but this August setup is still specifically K2.6. Do not rename a K2.6 response, benchmark, trial, or model ID as K3.

Five-minute evaluation API check

Create an NVIDIA developer account, open the Kimi K2.6 model page, and generate an API key if the trial is available to your account. Then test the documented OpenAI-compatible endpoint:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11

export NVIDIA_API_KEY="nvapi-your-key-here"

curl https://integrate.api.nvidia.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $NVIDIA_API_KEY" \
  -d '{
    "model": "moonshotai/kimi-k2.6",
    "messages": [{"role": "user", "content": "Return only: NIM K2.6 ready"}],
    "max_tokens": 128,
    "temperature": 0.6
  }'

Confirm the input is permitted under the restrictions above. Check the response model identity, HTTP status, token accounting, latency, and any signed-in quota before connecting a test agent. A successful trial request proves evaluation access at that moment; it does not prove an evergreen free entitlement or production permission.

OpenClaw trial/evaluation configuration

Use NVIDIA’s base URL and exact provider model ID only in an isolated, non-production evaluation profile supported by your OpenClaw version. Keep this profile within the input and data-use limits summarized above:

1
2
3

NVIDIA_API_KEY=nvapi-your-key-here
NVIDIA_BASE_URL=https://integrate.api.nvidia.com/v1
NVIDIA_EVALUATION_MODEL=moonshotai/kimi-k2.6

Do not enable this trial endpoint as an autonomous or production fallback. Before a bounded evaluation:

  1. Pin the exact model ID and record the checked date.
  2. Run a small read-only prompt and one tool-call test.
  3. Disable unattended execution and set token and retry limits in the surrounding agent.
  4. Check NVIDIA’s data, licensing, retention, and account-specific trial terms, including any API-service-specific disclosure.
  5. Remove the trial profile after evaluation.

NVIDIA’s trial endpoint is an internal compatibility-evaluation route. It is a separate product from Moonshot’s direct Kimi API, so provider behavior, retention, availability, and tool handling can differ even when the underlying model name is similar. See the Kimi access guide for the current listing and the limits above.

K2.6 is not K3

Kimi K3 is Moonshot’s latest flagship in this stack, with a 1M context window and released weights under the Kimi K3 License. NVIDIA’s hosted route in this August example was K2.6. Choose based on the actual requirement:

RequirementRoute to evaluate
NVIDIA-hosted OpenAI-compatible trialmoonshotai/kimi-k2.6 on NIM
Newest managed Kimi flagshipkimi-k3 through Moonshot
Self-hosted K3 evaluationOfficial K3 weights and license
Cheaper routine Kimi codingKimi K2.7 Code through Moonshot

The Kimi access guide keeps those plans, models, and prices separate. The Smart Spend guide explains why a trial route is not the same measure as accepted-result cost.

Sources and status

Related: Free Stack, Kimi K3, Kimi access, and OpenClaw provider policies.

NVIDIA API terms and the live catalog were rechecked October 3, 2026. The August K2.6 setup remains historical. Account access, payment and credits, quota, expiry, rate limits, and service-specific retention remain untested or unresolved.