Alibaba: the Qwen family of open models
Try Alibaba’s open Qwen models in a wide range of sizes, ready to fine-tune and deploy on Crusoe infrastructure.
Find the right size for your dataset
Fine-tune several Qwen model sizes on your own dataset with the same Serverless Fine-Tuning workflow, then compare quality, latency, and cost side by side. Pick the version that delivers what you need, judged on your own results rather than public benchmarks.
A specialist model for every task
Smaller Qwen models are well suited to narrow, high-volume jobs. Use Self-Serve Deployments to run separate fine-tuned variants for classification, extraction, or routing, each on its own dedicated endpoint, so every workload scales independently and one spike never slows down another.
Fine-tune and serve in one place
Deploy a fine-tuned Qwen checkpoint straight to its own endpoint without leaving the platform you trained on. With a dozen Qwen sizes available, you can retune at a different size and redeploy as your workload shifts, with no exporting, re-uploading, or moving between providers.
Fine-tune and deploy in Crusoe Intelligence Foundry
Frequently
asked questions
Alibaba Cloud, the cloud computing arm of Hangzhou-based Alibaba Group, develops the Qwen family and releases open models across a wide range of sizes and modalities, with Apache 2.0 licensing for the sizes most teams deploy. Qwen has become one of the most widely used bases for fine-tuning small and mid-size models. On Crusoe Cloud, Qwen models are available for Serverless Fine-Tuning, Self-Serve Deployments, and in some cases Serverless Inference, running on Crusoe-operated infrastructure with no customer data sent to Alibaba. Request a Tailored Deployment for bespoke optimization and a dedicated endpoint.
The best Qwen model depends on your priority: accuracy, speed, or cost. Qwen’s largest mixture-of-experts models score highest on reasoning, coding, and knowledge benchmarks while activating only a fraction of their parameters per token. Dense models use every parameter on every token, which makes them simpler to deploy and fine-tune. Small models are the fastest and cheapest to run, and they work well for high-volume tasks like classification, extraction, and routing.
Yes, Qwen’s open-weight models can run on any infrastructure, so prompts and outputs never need to pass through Alibaba’s own API or cloud. Data handling is governed by whoever hosts the model, not by the lab that trained it. On Crusoe Cloud, Qwen models run on Crusoe-operated infrastructure, and no customer data is sent to Alibaba.
Many Qwen models can reason step by step before answering, either through a thinking mode you toggle on and off or through separate thinking and instruct variants. Thinking mode improves results on math, coding, and multi-step problems, at the cost of more output tokens and higher latency. Non-thinking mode responds faster for chat, extraction, and simple tasks.
Qwen and Llama are both widely used open-weight families for fine-tuning, but they differ in licensing and range. Most Qwen models use the Apache 2.0 license with no user thresholds, while Meta Llama’s Community License adds conditions for very large products. Qwen also spans more sizes and architectures, including mixture-of-experts models, giving teams more options to match model size to task.
Qwen base models are pretrained on raw text and predict the next token, while instruct models are further trained to follow instructions and hold conversations. Instruct models are the right choice for chat, assistants, and most production applications. Base models suit teams doing extensive custom training, since they start without conversational habits that fine-tuning would need to override.