# Telnyx Inference

Telnyx Inference is a hosted API for open-weight models, running on GPU infrastructure Telnyx says it owns rather than rents. The interface is OpenAI-compatible, so it slots in wherever a base URL can be changed. Everything on this page comes from Telnyx's own release note announcing [[kimi-k3]] on the platform, dated 2026-07-28 — a vendor document, and worth reading as one.

## The catalog

Telnyx lists K3 (model ID `moonshotai/Kimi-K3`) alongside Kimi K2.6, GLM-5.2-FP8, and MiniMax M3. That is a catalog of Chinese open-weight models with no Western frontier models in it, which is the same shape [[crofai]] serves — and GLM-5.2-**FP8** carries the same tell, a model ID that discloses the serving quantization. Most providers don't put that in the name, and it matters: what you get from a hosted endpoint depends on how the weights were compressed for serving, not only on which model was published. See [[llm-quantization]] for why the format matters as much as the bit width. GLM-5.2 is [[z-ai]]'s model, and [[glm-5-2-step-change-for-open-agents]] covers why it was the first open-weight model to work as a general coding agent.

## What Telnyx claims about K3 on the platform

The release note advertises a **1M-token context window**, with the stated use cases being codebase analysis, long document processing, and multi-turn agent sessions; "stable long-context performance" is asserted rather than measured. **Native vision** — text, images, and video accepted by the same model, with no separate vision adapter. **Configurable reasoning effort** at three levels, low, high, and max, trading compute for reasoning depth per request. **Tool calling and structured output**, including function calling, dynamic tool loading, and JSON-schema-constrained output; the constrained-decoding claim is the one worth testing yourself, since [[structured-output-benchmark]] found a 15–30 point gap between producing valid JSON and producing *correct* values on every frontier model it measured. And **prompt caching by default** — automatic prefix caching for repeated prompt prefixes across requests, which is why the price sheet has a cached-input row.

## Pricing

| Token type | Price per 1M tokens |
| --- | --- |
| Cached input | $0.27 |
| Input | $2.70 |
| Output | $13.50 |

Cached input is exactly one tenth of uncached input, and output is 5× input. With caching on by default, an agent loop that keeps a stable system prompt and tool schema pays the $0.27 rate on most of its prefix — which is the main reason the three-tier sheet exists rather than a single input price.

## The framing, read as marketing

The note's "why it matters" section argues that "the competitive advantage in AI is shifting from who builds the smartest model to who builds the infrastructure that decides where every request runs" — a claim an inference provider has an obvious interest in making. It also says K3 "competes with closed-source frontier models from Anthropic and OpenAI" on coding, reasoning, and agentic knowledge work, without naming a benchmark or a number.

Two smaller wobbles are worth flagging. Telnyx calls K3 "the world's first open-source model in the 3-trillion-parameter class," which stretches 2.8T into a 3T class and uses "open-source" where "open-weight" is the accurate term — Moonshot published weights and modeling code under its own license, not source under an OSI license. Sebastian Raschka's architecture read in [[kimi-k3]] agrees on the substance (K3 is by a wide margin the largest open-weight model) without the rounding.

## Where it sits

For anyone deciding where to run K3, this is one of three published options within days of the weights dropping: a rented endpoint here, a managed endpoint at Modal (the one behind the anecdote in [[rl-finetune-beats-frontier]]), or [[deltafin]] on your own hardware at 14.6 seconds per token. The OpenAI-compatible surface is what makes those interchangeable at all — the same property [[cc-mirror]] leans on to point Claude Code at alternative providers.