Qwen3-VL-30B-A3B-Thinking

Qwen3‑VL‑30B‑A3B‑Thinking is a multimodal reasoning model from the Qwen3‑VL series, built on a Mixture‑of‑Experts (MoE) architecture. The model has a total of 30 billion parameters, of which only 3 billion are active during inference. It is based on a hybrid multimodal block that integrates the DeepStack visual encoder with the Qwen3‑LM language core. The connection between them is handled by the Interleaved‑MRoPE positioning mechanism, which precisely distributes frequency features across time, width, and height. This solution gives the model stable spatial perception (including 3D grounding) and efficient synchronization of video timestamps with textual semantics. As a result, it forms a unified visual‑textual stack in which the text transformer processes visual representations as n‑dimensional tokens seamlessly embedded into the common reasoning context.

Like all models in the series, Qwen3‑VL‑30B‑A3B‑Thinking supports a 256 K‑token context window (expandable to 1 M), allowing it to work with multi‑hour videos, large books, and complex agent pipelines without loss of data coherence. The “Thinking” (reasoning) mode provides an elevated cognitive level: the model generates responses step by step, demonstrating strict causal reasoning and well‑grounded conclusions.

The model is particularly effective in scenarios requiring the integration of textual and visual data with deep reasoning, such as the analysis of illustrated documents, semantic understanding and transcription of videos, OCR systems (supporting 32 languages), visual programming (code generation from visual sketches), and scientific or educational applications.

Thanks to vendor‑side quantization in FP8 mode, the model can be run with a practical context length on two RTX 4090 GPUs, making it accessible to a wide range of researchers and developers.

Announce Date: 26.09.2025
Parameters: 31B
Experts: 128
Activated at inference: 3B
Context: 263K
Layers: 48
Attention Type: Full Attention
Developer: Qwen
Transformers Version: 4.57.0.dev0
License: Apache 2.0

Use our pre-built public endpoints for free to test inference and explore Qwen3-VL-30B-A3B-Thinking 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 support@immers.cloud or send your request to the sales department at sale@immers.cloud.