Built for the workload that re-sends its prompt

Performance, and the conditions attached

We run two serving profiles and never mix them, because the settings that win on single-stream latency lose on aggregate throughput. They are separately priced, and every row below is tagged with the profile that produced it. Read the tag before you compare: the fastest number on this page and the cheapest price on this page are not available at the same time, and quoting them together would overstate what you actually get.

Speculative decoding runs in the fastest profile only: it raises per-request speed and lowers total throughput, so the batch profile does not use it. That is also why the fastest profile costs more — it genuinely burns more GPU per token, and its output price sits above every competitor's list price on this model. Take it only if latency is worth more to you than price. Everything else on this page — the cost table, the pricing section — quotes the default cheapest profile.

Measured on our own deployment on a single RTX 3090, 2026-08-21, reproducible on the same hardware. Read the throughput rows carefully: they used 128-token prompts. At a realistic 6,500-token agent prompt, prefill is the large majority of the GPU work, so aggregate throughput at that shape is materially lower than 873 tok/s. We are re-running the whole matrix at median shape and will republish it here; until then, do not compare this figure against a provider's live-traffic numbers, because they are not measuring the same thing. Independent per-provider throughput and latency are measured continuously by OpenRouter from real traffic — compare us there once we are listed, not here.

First-token time is a function of your prompt, not a constant

The ~100 ms above is a short prompt. Prefill runs at roughly 1,810 tokens per second, so a real agent prompt costs real time — unless its prefix is already cached. Most providers publish only the first column.

Change one line

When your key is issued, this is the whole integration. Point your existing OpenAI client at our endpoint — streaming and logprobs are supported, the latter of which only about a quarter of endpoints return at all.

# Works with the OpenAI SDK you already have installed.
client = OpenAI(
    base_url="https://api.jetinfer.com/v1",
    api_key="sk-...",
)

response = client.chat.completions.create(
    model="qwen3.8-27b",
    messages=[{"role": "user", "content": "Summarize this changelog."}],
    stream=True,
)

Pricing

Two profiles, separately priced. The cheapest profile is the default and is what every other number on this page is costed against; the fastest profile exists for buyers who value latency over price, and is deliberately priced above the market because it costs us more to serve.

Cheapest profile — default

Fastest profile — opt in

Ranges cover the providers serving this model class as of 2026-08-21 — the seven on OpenRouter for the default profile, and the 27B-class endpoints on HuggingFace for the fastest profile, which is a different field at different prices. Cached input bills at one tenth of the input rate on both profiles; a cache write costs the normal input price and nothing extra. The fastest profile trades aggregate throughput for single-stream speed, which is why it costs more per token. Prepaid credit only during early access, and you are never billed for a request we failed to serve.

Quantization

We serve int4 weights (W4A16), declared as "quantization": "int4" in our /v1/models response so it is machine-readable and filterable. Most endpoints on this model run fp8 or bf16; int4 is what makes the pricing possible, and you should know which one you are buying before you buy it.

Capacity, stated before you depend on it

This is early access on a small deployment, and pretending otherwise would only waste your integration time. Here is exactly what will stand behind the endpoint at launch.

What we are not offering: a contractual uptime SLA, or multi-region redundancy. Capacity gets added when demand justifies it, which is why early access is capped rather than open — an oversubscribed endpoint would wreck the measurements above, and those measurements are the whole pitch. If you need a committed SLA before you can build on us, say so in the form and we will tell you plainly whether we are there yet.

Data handling

Prompts and completions are never written to disk, logs, or analytics. We retain only operational metadata — request ID, key identifier, token counts, latency, and status — for billing and debugging, for 13 months. No customer data is used for training, fine-tuning, or evaluation. Inference runs on rented GPU capacity currently located in the Czech Republic. Full terms are in the privacy policy, and a data processing agreement is available on request.

Get on the early access list

The endpoint is not live yet — it is in final validation. Tell us what you intend to run on it and we will issue a key with trial credit as capacity comes online, capped, in the order requests arrive. Four fields, and a real person reads them.

The form opens your mail client with the answers filled in, so nothing is sent anywhere until you press send and you can see exactly what you are sending. Prefer to write it yourself? [email protected].