Qwen3.5-4B

Qwen3.5-4B is a small 4-billion-parameter model optimized for deployment on edge devices and mobile platforms. Its hybrid architecture includes 32 layers, 8 of which feature full attention, enabling efficient sequence processing with minimal computational cost. Despite its compact size, the model retains all the technical innovations of the Qwen3.5 series, including native multimodality and a 262K-token context window, allowing it to process long documents even on memory-constrained devices.

On benchmarks, the model delivers results surpassing many models twice its size. In language tests such as MMLU-Pro (79.1) and GPQA Diamond (76.2), it outperforms Qwen3-Next-80B-A3B-Thinking in several scenarios. In agentic tasks like TAU2-Bench (79.9), it achieves results comparable to models 20 times its size, confirming its effectiveness in planning and tool use. Its multimodal capabilities are also strong: scoring 85.1 on Mathvista (mini) — only slightly behind the 9B model — and achieving top-tier results on CountBench (96.3) and MMBench (89.4). This makes it ideal for tasks involving object, scene, and document recognition on memory-limited devices.

The model's uniqueness lies in bringing the qualities of "large" AI to the edge. It is an ideal solution for mobile applications, drones, robots, and smart cameras requiring fast, local analysis of visual and textual information without an internet connection. It stands out from other models in its class through its rare combination of deep multimodal capabilities and agentic "reasoning" in such a compact format.

Announce Date: 27.02.2026
Parameters: 5B
Context: 263K
Layers: 32, using full attention: 8
Attention Type: Linear Attention
Developer: Qwen
Transformers Version: 4.57.0.dev0
vLLM Version: 0.17.0
License: Apache 2.0

Use our pre-built public endpoints for free to test inference and explore Qwen3.5-4B capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.