Open Models for Every Modality, Built to Be Deployed
Qwen is the most complete open model family available today, and completeness is the point. Alibaba releases not one flagship but a full spectrum — dense models small enough to serve at volume alongside sparse models in the trillion-parameter class, plus dedicated vision, speech recognition, text-to-speech, and embedding models. They share one API surface, one set of conventions, and in most cases a permissive licence, which makes moving between them a configuration change rather than a rewrite. On DEVUP AI the whole family is reachable through a single OpenAI-compatible endpoint, billed in Algerian Dinar.
Most model developers ship a flagship and let everything else follow from it. Qwen ships a matrix. Within a single generation you will find a compact dense model that runs comfortably at production volume, a sparse Mixture-of-Experts model in the trillion-parameter class, and separate models built specifically for vision, speech recognition, speech synthesis, and embeddings. The result is a family where the right tool for a task usually already exists rather than needing to be approximated by a general model.
The architectural direction across recent generations is consistent: reduce the cost of long context rather than simply extend it. Recent models interleave gated linear attention with periodic full attention, typically three linear blocks for every full one. Most of the depth carries a constant recurrent state instead of a growing key-value cache, while regular full-attention layers restore exact retrieval where it matters. This is why models in this family handle very long inputs without the memory profile that length normally implies.
Thinking is unusually configurable here. Recent models expose three independent controls: whether the model reasons at all, how deeply it reasons across three levels, and whether reasoning carries forward across turns. That last setting is enabled by default, and for a reason worth understanding — preserving the reasoning trace keeps the conversation prefix stable, which improves both decision consistency in agent loops and prompt cache utilisation. Reasoning history is not purely a cost; on this family it is partly a saving.
One thing to check rather than assume: capability does not track size in this family the way intuition suggests. A compact model in a given generation may accept images and video while the far larger flagship in the same generation is text-only. Licences differ too, with some models permissively licensed and others carrying model-specific terms. Read the individual model page before assuming that the bigger number is the more capable choice.
Every Qwen model on DEVUP AI is reachable through the same OpenAI-compatible endpoint with one API key, billed in Algerian Dinar with local payment methods. Switching between them — including between text, vision, speech, and embedding models — is a change to the model field and the endpoint, not an integration.
Available Qwen Models
Deploy and access all published models in this family with unified DZD pricing and zero foreign card requirement.
| Model | Context | DZD in / 1M tokens | DZD out / 1M tokens | Actions |
|---|---|---|---|---|
| Qwen3.5-122B-A10BQwen/Qwen3.5-122B-A10B | 262K Tokens | 102 | 840 | View model |
| Qwen3.8-27BQwen/Qwen3.8-27B | 262K Tokens | 140 | 1050 | View model |
| Qwen3.8-2.4T-A95BQwen/Qwen3.8-2.4T-A95B | 262K Tokens | 700 | 2100 | View model |
| Qwen3.8-MaxQwen/Qwen3.8-Max | 256K Tokens | 578 | 1733 | View model |
| Qwen3.7-MaxQwen/Qwen3.7-Max | 256K Tokens | 2187.5 | 6475 | View model |
| Qwen3-30B-A3BQwen/Qwen3-30B-A3B | 40K Tokens | 39.2 | 164 | View model |
| Qwen3.5-397B-A17BQwen/Qwen3.5-397B-A17B | 262K Tokens | 157.5 | 1050 | View model |
| Qwen3-Embedding-0.6BQwen/Qwen3-Embedding-0.6B | 32K Tokens | 3.5 | — | View model |
| Qwen3-Embedding-4BQwen/Qwen3-Embedding-4B | 32K Tokens | 6.5333 | — | View model |
| Qwen3-Embedding-8BQwen/Qwen3-Embedding-8B | 32K Tokens | 3.5 | — | View model |
| Qwen3-32BQwen/Qwen3-32B | 40K Tokens | 28 | 98 | View model |
| Qwen3-Next-80B-A3B-InstructQwen/Qwen3-Next-80B-A3B-Instruct | 262K Tokens | 31.5 | 385 | View model |
| Qwen2.5-72B-InstructQwen/Qwen2.5-72B-Instruct | 32K Tokens | 126 | 140 | View model |
| Qwen3-14BQwen/Qwen3-14B | 40K Tokens | 42 | 84 | View model |
| Qwen3-Max-ThinkingQwen/Qwen3-Max-Thinking | 256K Tokens | 1050 | 5250 | View model |
| Qwen3.5-9BQwen/Qwen3.5-9B | 262K Tokens | 35 | 52.5 | View model |
| Qwen3.6-35B-A3BQwen/Qwen3.6-35B-A3B | 262K Tokens | 52.5 | 332.5 | View model |
| Qwen3.6-27BQwen/Qwen3.6-27B | 262K Tokens | 112 | 1120 | View model |
| Qwen3.5-35B-A3BQwen/Qwen3.5-35B-A3B | 262K Tokens | 49 | 350 | View model |
| Qwen3.5-27BQwen/Qwen3.5-27B | 262K Tokens | 91 | 910 | View model |
| Qwen3-VL-30B-A3B-InstructQwen/Qwen3-VL-30B-A3B-Instruct | 262K Tokens | 70 | 308 | View model |
| Qwen3-VL-235B-A22B-InstructQwen/Qwen3-VL-235B-A22B-Instruct | 262K Tokens | 70 | 308 | View model |
| Qwen3-MaxQwen/Qwen3-Max | 256K Tokens | 1050 | 5250 | View model |
Call any Qwen model
Every model on this page uses the same API key and standard OpenAI-compatible endpoint. The example below uses the featured model Qwen/Qwen3.8-27B.
from openai import OpenAI
import os
client = OpenAI(
api_key=os.environ["DEVUP_API_KEY"],
base_url="https://api.devupai.com/v1"
)
response = client.chat.completions.create(
model="Qwen/Qwen3.8-27B",
messages=[
{"role": "user", "content": "Hello world!"}
],
max_tokens=1024
)
print(response.choices[0].message.content)import OpenAI from "openai";
const client = new OpenAI({
apiKey: process.env.DEVUP_API_KEY,
baseURL: "https://api.devupai.com/v1",
});
async function main() {
const response = await client.chat.completions.create({
model: "Qwen/Qwen3.8-27B",
messages: [
{ role: "user", content: "Hello world!" },
],
max_tokens: 1024,
});
console.log(response.choices[0]?.message?.content);
}
main();curl -X POST "https://api.devupai.com/v1/chat/completions" \
-H "Authorization: Bearer $DEVUP_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "Qwen/Qwen3.8-27B",
"messages": [
{ "role": "user", "content": "Hello world!" }
],
"max_tokens": 1024
}'Frequently Asked Questions
Start with a compact dense model from the current generation at medium reasoning effort. It handles most production work, serves predictably under concurrency, and in many workloads lands within a few points of the flagship. Move up when your task is knowledge-heavy or involves long-horizon agent execution. Check the model page first — the compact model may support inputs the flagship does not.
No, and this catches people out. Within the same generation, a compact model may accept images and video while the much larger flagship is text-only. On general software engineering and instruction following the two can land within a few points of each other. Size buys hard knowledge and long-horizon capability; it does not automatically buy modality support or a non-thinking mode.
The first switches reasoning on or off entirely — turn it off for classification, routing, and formatting. The second sets depth across three levels, and the default is the deepest one, so set it explicitly. The third controls whether reasoning carries forward across turns; it is on by default and generally should stay on for multi-turn and agent work, since a stable conversation prefix improves both consistency and cache hit rates.
Per turn, yes. Per task, not necessarily. Qwen notes that on multi-turn agentic work, shallower analysis produces more failures and more retries, which can raise total latency and token consumption above what a deeper single pass would have cost. Measure the whole task, not the individual call.
Some do, and the family includes models built specifically for each. Vision-language models accept images and video; separate models cover speech recognition, speech synthesis, and text embeddings. Support is per model, not per family — always check the model page rather than assuming a shared capability.
Most are, but not all. Some models in the family carry model-specific licence terms rather than a standard permissive licence. If you are building something commercial, check the licence on the individual model page before committing to it.
Not for the core call. Text and vision-language models share the chat completions endpoint; embedding models use the embeddings endpoint. Recommended sampling parameters differ per model and per thinking mode, so read the model page for those. Switching models is a change to one field, not a re-integration.
