Qwen models

The family

Qwen is Alibaba's open-weight line, published under Apache 2.0 and released as whole checkpoints rather than API access. It is the widest family in our catalogue: dense text and vision models in the size range batch work actually pays for, an omni checkpoint that reads audio and video in the same pass as text, and a set of embedding towers that put several modalities into one vector space.

Models that answer in text

qwen3.8-27b

The newest 27B dense checkpoint, and the only model here that reads video without also carrying an audio tower. Its vision encoder has the smallest measured per-image overhead of the three image models we serve, and video is charged per sampled frame at one frame per second, up to a two-minute clip. It emits a reasoning trace, which a job can cap with a reasoning budget.

qwen3.6-27b

The previous 27B dense checkpoint: text and images, no video. It still scores higher than its successor on knowledge and general reasoning (see below), so it is not simply the older option - it is the broader one. Also a reasoning model.

qwen3-omni-30b

A mixture-of-experts checkpoint with 30B total parameters and about 3B active per token, carrying image, audio and video encoders at once. It is the only model here that reads audio: up to 40 minutes of clip, priced by duration through its own encoder. It answers in plain text with no reasoning trace.

qwen3-14b

Text only, and the cheapest chat model we serve. A 14B dense reasoning model is enough for classification, extraction, routing and rewriting at volume, which is most of what a batch queue actually carries.

Models that answer with a vector

The bracketed range is what you may ask for with the dimensions field; the first number is what the model returns if you ask for nothing. The two qwen3-vl-embedding-* models put text, images and video into one vector space, so a text query can retrieve a page image or a clip without a second index. See embedding jobs for the request shape and pricing for the rates.

Benchmark scores

Published figures, each from the source named beside it. They are not comparable across rows: different benchmarks, different harnesses, and in most cases the model's own publisher doing the measuring. Use them to tell these models apart from each other, not to rank them against something scored elsewhere.

Two things worth knowing before you read too much into the table. The qwen3-14b figures are its thinking-mode scores; with thinking off the same model reads 54.8 on GPQA Diamond and 29.0 on LiveCodeBench, so which mode you run is a bigger lever than which of these models you pick. And the newer 27B does not beat the older one everywhere: 3.8 is tuned for agentic execution and coding, 3.6 is broader on knowledge.

Which one to pick

Simple classification, extraction, routing, rewriting

qwen3-14b. The cheapest per token and the cheapest minimum job price. If the task is more nuanced, or the rows carry images, go for qwen3.8-27b instead.

Documents, screenshots, product photos

qwen3.8-27b for its cheaper image encoding, or qwen3.6-27b if the work leans on general knowledge rather than on reading the page.

Video

qwen3.8-27b if the clip is silent, or qwen3-omni-30b if the soundtrack carries meaning.

Audio, transcription, call analytics

qwen3-omni-30b. It is the only model we serve with an audio encoder.

Search and retrieval indexes

qwen3-embedding-8b for text alone; qwen3-vl-embedding-8b when pages, images or clips have to sit in the same index as the text; qwen3-vl-embedding-2b when the corpus is large enough that a quarter of the price matters more than the last few points of recall.

Pass the model id exactly as it is written above when you submit a job. Rates for every model are on the pricing page, and every job is estimated and priced before it runs, so trying a second model costs you an estimate rather than a bill.