Models

A sparse mixture-of-experts model from Alibaba's Qwen team, and the reason a flat rate is possible at all.

What 35B-A3B actually means

The two numbers are total parameters and active parameters. Qwen 3.6 holds 35 billion weights but routes each token through only about 3 billion of them, picking a handful of specialised experts per token instead of running the whole network.

You get the knowledge of a 35B model at roughly the inference cost of a 3B one. Dense models make you pay for every parameter on every token; sparse ones don't. That gap is what a flat-rate plan is built on — a provider can only stop metering you if serving you doesn't scale linearly with how much you use it.

The 256K context window

262,144 tokens is roughly 200,000 words, or a mid-sized codebase. You can keep a long agent session, a stack trace, the files it touches and the conversation that led there all in context without pruning.

Long context is exactly what makes per-token pricing painful, because every turn resends everything before it. On a flat plan there's no reason to trim context to save money — the cost calculator shows what that's worth at your volume.

A larger mixture-of-experts model tuned for throughput, with a million-token window for whole-repository work.

What a 1M window is for

Four times the context of Qwen 3.6, which changes what you can attempt rather than just how long a session lasts. A million tokens holds an entire repository plus its history, so an agent can reason across modules it was never pointed at — the kind of task where a smaller window forces you to guess which files matter before you've understood the problem.

It's the same trade every MoE makes, at a larger scale: far more total knowledge, only a slice of it active per token.

Model upgrades are included

When a better open model ships, it goes into the catalogue at no extra cost and with nothing for you to do. The API doesn't change, so your integration keeps working as-is. You are buying capacity on the best open weights available, not a specific model that slowly ages.

This is the practical advantage of open weights, and it compounds with the other one: because we run them ourselves on European hardware, zero logs and EU-only processing are structural rather than a clause in a contract. No third party sits in the path to log anything.

Calling a model

Point any OpenAI-compatible client at the endpoint and name the model:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://run.llmax.ai/v1",
    api_key=os.environ["LLMAX_API_KEY"],
)

resp = client.chat.completions.create(
    model="qwen3.6",
    messages=[{"role": "user", "content": "Refactor this function to use async/await."}],
)

A GET /v1/models returns the authoritative list of what your key can reach. Setup for editors and chat apps — Cline, Continue, Aider, Open WebUI — is in the quickstart.