AI Models
What Is Kimi K3 and How Much Does the API Cost?
Kimi K3 is Moonshot AI's 2.8-trillion-parameter open model now available through the Telnyx Inference API at $2.70 per million input tokens and $13.50 per million output tokens as of July 2026. The model weights are freely downloadable from HuggingFace, but commercial users face license restrictions above $20 million in annual revenue.
What matters
- Kimi K3 costs $2.70/MTok input and $13.50/MTok output via Telnyx, close to Claude Sonnet 4.5's standard $3/$15 but pricier on output than GPT-4o at $10/MTok.
- The 2.8T-parameter model runs a 1M-token context window with native vision and an OpenAI-compatible API endpoint on Telnyx.
- Model weights are free on HuggingFace (99K downloads, no gating), but a custom license restricts commercial API services above $20M annual revenue.
- Cached input at $0.27/MTok is roughly one-fifth of GPT-4o's $1.25, making Kimi K3 attractive for repeated-prompt workloads like customer support bots.
What Is Kimi K3?
Kimi K3 is a large language model developed by Moonshot AI, a Beijing-based company behind the Kimi chatbot popular in China. Released in June 2026 and made available on the Telnyx Inference API on July 28, 2026, it uses a Mixture-of-Experts architecture with 2.8 trillion total parameters. The model handles both text and images with a 1-million-token context window, using techniques Moonshot calls Kimi Delta Attention and Attention Residuals for long-context performance.
On HuggingFace, Kimi K3 has accumulated 99,214 downloads and 7,991 likes as of late July 2026. It is not gated: anyone can download the weights without approval. Quantized 8-bit versions are also available for teams with hardware constraints.
Kimi K3 API Pricing on Telnyx
Telnyx, a cloud communications provider that also runs an AI inference platform, added Kimi K3 on July 28, 2026. The API uses model ID moonshotai/Kimi-K3 and is compatible with the OpenAI chat completions format. Verified pricing as of July 29, 2026:
- Cached input tokens: $0.27 per million
- Standard input tokens: $2.70 per million
- Output tokens: $13.50 per million
Telnyx does not currently split Kimi K3 pricing into separate short-context and long-context tiers. The same rates apply regardless of prompt length, up to the 1M-token context limit.
Kimi K3 vs GPT-4o vs Claude: Pricing Compared
Kimi K3 sits in the middle of the commercial API pricing range. Its input rate ($2.70) is slightly above GPT-4o ($2.50) and slightly below Claude Sonnet 4.5 ($3.00 at standard pricing from September 2026). Output pricing ($13.50) is higher than GPT-4o ($10.00) but lower than Claude Sonnet 4.5 ($15.00) and GPT-5.6-terra ($15.00).
The standout number is cached input. At $0.27 per million tokens, Kimi K3 costs roughly one-fifth of GPT-4o's cached input rate of $1.25. It is comparable to GPT-5.6-terra at $0.25 and Anthropic's prompt-caching read rate of $0.30. For applications that reuse the same system prompts or context blocks across many requests, such as customer support bots or content moderation pipelines, this rate difference compounds into meaningful monthly savings.
A note on Claude Sonnet 4.5: Anthropic currently runs introductory pricing of $2 input and $10 output per million tokens through August 31, 2026, making it cheaper than Kimi K3 on both dimensions until September.
Is Kimi K3 Actually Open Source?
Moonshot AI describes Kimi K3 as open source, and the model weights are indeed freely downloadable on HuggingFace with no gating. However, the license reviewed on July 29, 2026 includes commercial restrictions that standard open-source licenses (MIT, Apache 2.0) do not impose.
The key terms:
- You may use, copy, modify, distribute, fine-tune, and sell the model for any purpose.
- If you operate a Model-as-a-Service business with aggregate revenue above $20 million in any 12-month period, you must enter a separate commercial agreement with Moonshot AI.
- If your product exceeds 100 million monthly active users or $20 million in monthly revenue, you must display "Kimi K3" prominently in the user interface.
- Internal use that does not expose the model to third parties is fully exempt from these requirements.
This is best described as a source-available license with a revenue cap, not an OSI-approved open-source license. For startups and small businesses, it is functionally unrestricted. For enterprises and API resellers, it requires a direct deal with Moonshot AI.
Who Should Use Kimi K3?
Kimi K3 makes the most sense for three groups:
- Developers optimizing for cost on high-volume workloads. The cached input pricing undercuts every major Western provider. Combined with Telnyx's OpenAI-compatible API, switching requires minimal code changes.
- Teams with existing GPU infrastructure. Download the weights from HuggingFace and run inference locally to avoid per-token API costs entirely. Quantized versions reduce the hardware requirement significantly.
- Companies building multi-model pipelines. Adding Kimi K3 alongside GPT-4o or Claude provides pricing leverage and a fallback provider without vendor lock-in.
It is less suitable for enterprises above the $20M revenue threshold that need guaranteed license compliance without negotiating a separate deal, or teams that depend on provider-managed safety tooling and SLAs at the level Anthropic and OpenAI offer.
At a glance
| Model | Input | Cached Input | Output |
|---|---|---|---|
| Kimi K3 (Telnyx) | $2.70 | $0.27 | $13.50 |
| GPT-4o (OpenAI) | $2.50 | $1.25 | $10.00 |
| Claude Sonnet 4.5 (Anthropic) | $3.00 | $0.30 | $15.00 |
| GPT-5.6-terra (OpenAI) | $2.50 | $0.25 | $15.00 |
FAQ
Is Kimi K3 free to use?
The model weights are free to download from HuggingFace and run on your own hardware. Using it through the Telnyx API costs $2.70 per million input tokens and $13.50 per million output tokens. The license allows free commercial use up to $20 million in annual revenue; beyond that, a separate agreement with Moonshot AI is required.
How does Kimi K3 compare to GPT-4o?
Kimi K3 is larger (2.8T parameters vs GPT-4o's undisclosed count) with a longer context window (1M vs 128K tokens). API pricing is slightly higher on input ($2.70 vs $2.50 per MTok) and output ($13.50 vs $10.00). The biggest difference is cached input: Kimi K3 charges $0.27 per MTok vs GPT-4o's $1.25, a 4.6x savings for repeated-prompt workloads.
Can I run Kimi K3 on my own server?
Yes. The model weights are not gated on HuggingFace and are available under a permissive custom license. The full 2.8T-parameter model requires substantial GPU memory, though 8-bit quantized versions reduce hardware needs. Internal use that does not expose the model to third parties is exempt from all commercial license restrictions.
Related reading
best local LLM tools, Claude tool profile, LM Studio Bionic, radar