Qwen (Alibaba): open models from 27 billion to 2.4 trillion parameters

Contents

1. What it does for a company
2. Strengths and limits
3. Pricing
4. Questions to ask before choosing Qwen
5. Sources

Type	Open-weight model family
Pricing	Published
Website	huggingface.co

Alibaba’s Qwen3.8 generation runs from a dense 27-billion-parameter model that a single server can host to a 2.4-trillion-parameter flagship, and the two come under very different licences⁠12Qwen/Qwen3.8-27BQwen (Hugging Face)Qwen3.8-Max LicenseQwen (Hugging Face).

Best for: The widest range of sizes, from a 27-billion-parameter Apache model to a 2.4-trillion-parameter flagship.

Not the right fit for: Service businesses above $50 million in revenue that want to resell the flagship without a separate licence.

What it does for a company

Qwen3.8-27B is the practical entry point. It is a dense model that reads text, images and video, with a 262,000-token context, released under Apache 2.0⁠1Qwen/Qwen3.8-27BQwen (Hugging Face). Its vendor-reported SWE-bench Pro score is 61.7⁠1Qwen/Qwen3.8-27BQwen (Hugging Face).

The flagship, Qwen3.8-2.4T-A95B, is a text-only mixture-of-experts model with 2.4 trillion parameters, 95 billion active per token, and a native context of 262,144 tokens that can be extended to about one million. Qwen documents OpenAI-compatible serving through vLLM and SGLang⁠3Qwen/Qwen3.8-2.4T-A95BQwen (Hugging Face). Vendor-reported results include 92.6 on GPQA Diamond and 67.7 on SWE-bench Pro⁠3Qwen/Qwen3.8-2.4T-A95BQwen (Hugging Face).

The flagship’s licence is the point to read carefully. Above 100 million monthly users or $20 million in monthly revenue, the product must display the model’s name. Businesses that offer models as a service or AI work assistants and earn more than $50 million need a separate licence; internal use inside a company is exempt⁠2Qwen3.8-Max LicenseQwen (Hugging Face). Neither model card publishes a fine-tuning guide or hardware guidance.

Strengths and limits

Strengths

• The 27B model is Apache 2.0 and multimodal, a practical size for a single server.
• OpenAI-compatible serving through vLLM and SGLang is documented.
• Vendor-reported SWE-bench Pro of 61.7 for the 27B model.

Limits

• The flagship’s licence requires showing the model name above 100 million monthly users or $20 million monthly revenue.
• Model-as-a-service and AI-assistant businesses above $50 million revenue need a separate licence for the flagship; internal use is exempt.
• Fine-tuning documentation is not on the model cards.

Choose another option if: You want a plain MIT or Apache licence at flagship scale (DeepSeek) or a European developer (Mistral).

Pricing

Weights are free. Together AI lists Qwen3.7-Plus at $0.32 per million input tokens and $1.28 per million output tokens.

Questions to ask before choosing Qwen

• Is our use internal, or will we offer the model to customers, which brings the flagship’s thresholds into play?
• Does the Apache-licensed 27B model meet our quality bar on our own tasks?
• Does our security policy permit weights from a Chinese developer on hardware we control?
• How will we fine-tune and evaluate it, given that we found no official guide?

See how Qwen compares with the other options in the main guide.

Sources

1. Qwen/Qwen3.8-27B. Qwen (Hugging Face). Accessed 9 Nov 2026.
2. Qwen3.8-Max License. Qwen (Hugging Face). Accessed 9 Nov 2026.
3. Qwen/Qwen3.8-2.4T-A95B. Qwen (Hugging Face). Accessed 9 Nov 2026.
4. Pricing: serverless inference, dedicated inference and GPU clusters. Together AI. Accessed 9 Nov 2026.

Further reading

• Buyer's guide · Updated quarterlyThe best open-source AI in 2027, for companies that want to own itAn analysis of the open-weight model families a company can run today and the layer that turns one into its own, compared on seven published criteria, with sourced licences, costs and the established path from rented AI to owned AI.
• Licence guideOpen source vs open weights: what each licence lets a company doAlmost every model sold as open-source AI publishes its weights but not its training data. The difference rarely matters for internal use. It matters a great deal for companies that resell a model or build a product on it.
• Practical guideHow to train your own AI model on company dataMost of the work in training a company model happens before any training starts. A step-by-step account of the data work, the choice of base model, the training methods and the rules that apply in Europe.

Also in: Deutsch · Français · Español · Italiano · Nederlands · Polski