Telnyx said in a July 28 release note that Moonshot AI's Kimi K3 is available through the Telnyx Inference API. Telnyx is led by co-founder and CEO David Casem.
The release advances a strategy Casem has been building toward from an unusual starting point. A self-taught engineer who studied economics at Swarthmore College, Casem founded Telnyx in 2009 and built its first phone system with open-source software and an FCC license. Telnyx spent the following years turning carrier infrastructure, phone numbers and communications networks into APIs. Casem is applying the same playbook to AI inference: own the expensive infrastructure, hide its operational complexity and charge developers for what they use.
Kimi K3 comes from Moonshot AI, founded in early 2023 and led by Yang Zhilin. Yang earned a computer science doctorate from Carnegie Mellon University and co-authored the influential Transformer-XL and XLNet papers. That research history matters here: Moonshot AI has built its product identity around long-context models, and K3 extends that thesis to a one-million-token context window.
Telnyx's July 28 note says K3 is already available through its hosted API. Moonshot AI's Kimi K3 materials describe 2.8 trillion total parameters, 104 billion active parameters and a mixture-of-experts architecture, while the official Kimi K3 page on Hugging Face positions K3 as an open-weights release.
Frontier-scale inference without the cluster
Kimi K3's headline parameter count overstates how much of the model runs for each token, since only 104 billion parameters are active at a time. The total size still creates a demanding serving problem.
Telnyx is selling access to that machinery through an OpenAI-compatible inference API. Developers select moonshotai/Kimi-K3
, submit requests through a familiar chat-completions interface and leave GPU provisioning to Telnyx. Telnyx also points developers to its Inference documentation for implementation details. Telnyx's release note says the API supports function calling, dynamic tool , JSON-schema-constrained output and automatic prompt caching. Reasoning effort can be set to low, high or max.
The model accepts text, image and video inputs, while its context window reaches 1,048,576 tokens. Those specifications make Kimi K3 relevant for long code repositories, document collections and agent sessions that need to retain substantial working context. They do not guarantee useful results across that entire window.
Telnyx charges $0.27 per million cached input tokens, $2.70 per million uncached input tokens and $13.50 per million output tokens. Prompt caching could materially change the bill for agents that repeatedly send the same system instructions, tool definitions or repository context. Output remains the expensive side of the request, giving developers an incentive to control response length and agent loops.
Telnyx says Kimi K3 competes with proprietary frontier models across coding, reasoning and agentic work. That is still a vendor benchmark claim rather than proof that K3 will match closed models in a given production workflow, so teams should test latency, tool use and output quality on their own workloads before moving traffic.
Casem's infrastructure bet
Kimi K3 reached hosted inference services quickly after its July rollout. Fireworks AI has also published its own Kimi K3 hosting pitch, and inference providers can package the model, optimize the serving stack and compete on price, speed, data controls and reliability.
Telnyx differentiates through the infrastructure surrounding the model. Telnyx operates communications APIs, phone numbers, speech-to-text, text-to-speech, messaging and voice-agent tools alongside inference. A developer building a phone-based agent could use Telnyx for the call, transcription, model request and generated voice instead of stitching together several vendors.
Telnyx lays out that position in a Telnyx essay on AI connectivity, and its broader pitch is control of the stack from carrier network to AI inference. That could matter for real-time applications such as voice agents, though Telnyx has not published K3 latency or throughput measurements in the release note.
Casem gets a frontier-scale model that can pull inference traffic into Telnyx's wider stack. Each side contributes capital-intensive work: Moonshot AI trained the model, while Telnyx operates the network and GPUs that turn it into a production endpoint.
The term "open-source" needs precision in this case. Moonshot AI has released K3 under a custom Kimi K3 License. Some model-as-a-service businesses above $20 million in revenue must sign a separate Moonshot agreement before commercial use. The custom terms make open-weight the more precise designation for K3.
Telnyx's move shows how quickly a frontier model can become infrastructure inventory. Moonshot AI built Kimi K3 as an open-weight model, and hosting providers are moving to make it callable through standard APIs. Casem's bet is that the winning endpoint will be the one attached to the rest of an agent's operating stack, from inference to the phone network carrying the conversation.