Explore additional models
Frequently
asked questions
Qwen3.5 4B is good for high-volume classification, routing, and extraction after fine-tuning, where a small model needs to handle long inputs at very low latency. It is a dense 4B model with a 262,144-token context, text and image input, and support for 201 languages, released under the Apache 2.0 license. It is available for Serverless Fine-Tuning in Crusoe Intelligence Foundry.
Qwen3.5 4B is the lighter option. It needs less GPU memory and responds faster, which suits latency-sensitive or high-volume workloads. Qwen3.5 9B has more than twice the parameters and generally produces higher-quality output on complex tasks. Both share the Qwen3.5 architecture, context window, and language support, so the choice comes down to speed and cost versus quality. On Crusoe, Qwen3.5 4B also accepts image input.
Qwen3.5 4B is available for Serverless Fine-Tuning and billed per 1M training tokens, where training tokens equal dataset tokens multiplied by epochs. There are no GPU clusters to provision and no minimum commitment, and you can estimate your cost before submitting a job. Current rates are shown in the pricing section above.
Qwen3.5 4B delivers strong quality for its size, with the same architecture and language coverage as the larger Qwen3.5 9B at roughly half the memory. Fine-tuned small models often match much larger general models on a narrow task, which makes Qwen3.5 4B a common choice for production classifiers, routers, and extractors.
Yes. Qwen3.5 4B is available for Serverless Fine-Tuning. Upload a JSONL or Parquet dataset, run a LoRA job with pre-configured defaults, then download the weights under Apache 2.0 to deploy where you choose. Its small footprint makes training runs quick, so you can iterate on your data fast.
Yes, Qwen3.5 4B accepts both text and image input and returns text, so you can fine-tune it for visual tasks like document classification, receipt and form extraction, and screenshot routing. At 4B, it is one of the smallest multimodal models on Crusoe, suited to high-volume visual workloads where a larger vision model would be too slow or costly.


%204.png)
