Kimi K3, Moonshot AI’s 2.8 trillion parameter open model, is available on Telnyx Inference APIs. It’s called via an OpenAI-compatible endpoint, costs $2.70 per million input tokens and $13.50 per million output tokens, and supports one million token context with native vision and function calling.

This is the first concrete signal that Chinese open models aren’t confined to weight downloads on Hugging Face. Telnyx is a US infrastructure provider with its own GPUs: anyone building AI products in the US can call Kimi K3 with the same API request they’d use for GPT or Claude, without deploying local hardware or touching Chinese repositories.

As we reported on July 27, Kimi K3’s release triggered alarms on Wall Street and pushed Washington to evaluate targeted bans on Chinese open models. API availability shifts the terms of debate: as long as weights are just downloadable, using them stays a political and individual choice. When a US provider serves them as a metered product, choice becomes economic, and the calculation shifts to costs and performance, not geopolitics.

List prices are competitive against frontier models, but cost-per-token is only the start of the calculation. Latency, service stability, and quality on your specific tasks weigh as much as pricing. To decide if it’s worth it, the method doesn’t change: same task, same test cases, table of results.

In detail

Kimi K3 is a Mixture of Experts (MoE): a model with 2.8 trillion total parameters, of which only 104 billion activate per token. In practice, the model is massive but each individual request uses only a fraction of its capacity. This lowers inference cost versus a dense model of the same size, because fewer active parameters means less compute per token.

The one million token window lets you load entire codebases or very long documents in a single call. The model has native vision (accepts text, images, and video without a separate module) and supports function calling, JSON-structured output, and three configurable reasoning effort levels: low, high, max. That last point means you can decide case-by-case how much compute to dedicate to the response, trading depth against cost.

The fact that Telnyx serves it on its own GPUs is the interesting part. Until now, using a Chinese open model meant downloading weights and running inference locally, or relying on providers like Together.ai or Baseten. Telnyx isn’t an AI lab: it’s a telecommunications and network infrastructure company that invested in GPUs to offer inference. Its catalog already includes Kimi K2.6, GLM-5.2-FP8, and MiniMax M3, all open Chinese models. The move signals that open model inference is becoming a business for whoever has the hardware, not just for whoever built the model.

On pricing: $2.70 per million input tokens and $13.50 per million output tokens, with automatic prompt caching at $0.27 per million on repeated prefixes (Telnyx data, July 28, 2026). For comparison, frontier models from Anthropic and OpenAI cost more per token, but list price is only the beginning of the calculation. Caching helps with long, repetitive prompts typical of agentic workflows, but the real question is how much it costs to produce usable output, not how much a token costs. That depends on your use case and only measures true with a test bench.

The regulatory context weighs on the picture. On July 26, the White House chose the path of targeted bans on individual Chinese open models, avoiding a blanket prohibition. Moonshot had also been accused of distilling Anthropic’s Fable model. A US provider serving Kimi K3 via API sits in an operational gray zone: if Washington bans a specific model tomorrow, Telnyx will have to remove it from its catalog, and anyone who built on it faces a rewrite. It’s the concrete risk that comes with anchoring your supply chain to models outside Western lab control.

What remains unknown: how stable Kimi K3 on Telnyx is under load, what real-world latencies look like, and how it behaves on agentic tasks versus frontier models. Benchmarks published by Moonshot say it competes with Claude and GPT on coding and reasoning, but corporate benchmarks should be read as qualified marketing, not gospel. The only way to know if it holds up on your work is to test it on your own data.