Kimi K2.6: open weights, API costs and licence limits
Scoped recheck: 24 October 2026. Official API retirement notice, prices, model card and licence checked. No model download, API call or account test.
Kimi K2.6 from Moonshot AI is a multimodal mixture-of-experts model with open weights. Its card lists 1 trillion total parameters, 32 billion active parameters and 256K context. That does not make hosted API calls, the Kimi app or every Agent Swarm workflow free.
Official publisher logo hosted by its publisher. Official source. Independent guide, not affiliated or sponsored. Artwork and trademarks belong to their owners.
What do the parameter counts mean?
A mixture-of-experts model uses selected expert parts for each token. The card lists 1T total parameters and 32B activated, with a 400M-parameter MoonViT vision encoder.
Active parameters do not mean you only need memory for 32B parameters. Deployment still needs the full model arrangement and serving resources. No laptop-size hardware promise is established here.
Current context and modalities
The model card lists 256K context, image and video input, and thinking/instant modes. It marks video chat as experimental and currently supported only in the official API, not third-party deployments. The current API model list describes visual and text input. This does not establish audio/video generation or production support for every media type. Check the exact endpoint before depending on it.
Multimodal does not automatically mean it accepts every media type or generates audio and video. Use the exact endpoint and card instead of transferring another Kimi model's capabilities.
Open weights versus the Kimi product
The official product page describes Agent Swarm, websites, documents, slides, spreadsheets and Document-to-Skills. The model card advertises long-horizon execution and coordinated agents.
Those are model and product claims, not proof that a bare weight download includes a hosted app, persistent computer, integrations or a ready-to-use swarm. The surrounding software and permissions matter.
Read the Modified MIT licence
The retrieved Modified MIT licence permits use, modification and distribution under its conditions, including keeping the copyright and permission notice. If a commercial product or service using the software or derivatives has more than 100 million monthly active users or more than US $20 million monthly revenue, it requires prominently displaying "Kimi K2.6" in the interface. The original US-dollar threshold controls; this recheck does not supply a current rupee conversion. Read the full licence for your use.
If a commercial product or service using the software or derivatives has more than 100 million monthly active users or more than US$20 million monthly revenue, the licence requires prominently displaying "Kimi K2.6" in its interface.
The US$20 million threshold is approximately Rs 192.94 crore at the checked Rs 96.47 per dollar reference, but the licence's wording controls, not this illustrative conversion. Read the full terms for your actual use.
Current K2.6 API prices
The current rendered official pricing table still lists US $0.95 input, $4 output and $0.16 cache-hit input per million tokens for kimi-k2.6. Prices exclude applicable tax. Cached input is not every input. API inference is pay-as-you-go, separate from Kimi membership and Kimi Code. Extracted document text sent to the model is billed as input. No API call or account bill was tested. Older rupee illustrations are removed rather than presented as a current exchange rate.
Cached input is not every input. The pricing explanation says both input and output are billed, including extracted document text later sent to the model. Service tools, terms, tax and actual account billing need review.
Do not call it the latest model now
The current API model list includes K3.1 Pro, K3 and K2.7 Code as well as K2.6. Both the official model list and pricing docs announce that K2-series models will not be available to newly registered users from 31 October 2026; existing users are not affected by that October cutoff. All K2-series models are planned to retire by the end of December 2026. This is a hosted API transition notice, not a claim that the downloaded weights or licence disappear. Avoid starting a new dependency without a migration plan.
The older K2.6 product page still calls it the latest model and says it is free to use with paid advanced plans. That page does not establish an Indian account’s current free quota or remove API charges and retirement limits. Use the current API docs for service availability.
API prices and product access can change. Open weights can remain useful even after newer models appear, but comparison should be made against the actual task and cost.
Benchmark and autonomy boundaries
Moonshot's card publishes coding, reasoning, vision and agent results. These are developer-reported evaluations, not independent tests performed for this guide.
Claims about 24/7 background operation do not authorise an agent to send messages, spend money or change accounts without your approval. Long-running work needs budgets, logs, stop controls and narrow permissions.
A low-risk evaluation workflow
- Choose API access, the Kimi app or self-hosting deliberately.
- Read the exact current licence and service terms.
- For API use, confirm K2.6 and a small spending limit.
- For self-hosting, check memory and runtime requirements.
- Begin with public data and a task with a known answer.
- Inspect generated code, documents and source links.
- Keep tools read-only until they are needed.
- Review recipients, amounts and final changes before effects.
- Check actual usage and stop incomplete runs cleanly.
India, privacy and free verdict
The weights are available under the retrieved licence, but no Indian account, payment route or hosted free quota was verified. App subscriptions and Agent Swarm limits were not established from a logged-in account.
No model was downloaded, API key connected or autonomous task executed here. Malayalam quality, Indian data residency and self-hosted speed remain untested. Review cloud data handling before sending private material.
What changed since the original post?
The 24 October scoped recheck adds the 31 October new-user cutoff and planned end-December K2-series retirement, current rendered K2.6 prices before tax, and the model card’s experimental official-API-only video limitation. It preserves the weights-versus-hosted-service distinction and exact Modified MIT attribution condition. No model, account, billing or Agent Swarm workflow was tested.
Frequently asked questions
Are the weights ordinary MIT?
The licence is Modified MIT with an additional large-commercial-use attribution condition.
Is the API free because weights are open?
No. The current API page lists paid token rates.
Was Agent Swarm tested?
No. Product access and task limits were not tested.
Sources: Official K2.6 model card · Exact Modified MIT licence · Official K2.6 product page · Current official API models and prices · Official pricing explanation · Approximate exchange-rate reference. Checked 7 October 2026.
Scoped recheck: current model list and retirement notice · rendered prices, tax and retirement notice · model card and experimental video limits · exact licence. Checked 24 October 2026. Hosted access and self-hosting were not tested.