Qwen3-VL 30B-A3B Instruct

On October 4, 2025, Alibaba Cloud's Tongyi Qianwen open-sourced the Qwen3-VL-30B-A3B-Instruct and Thinking models alongside FP8 versions and the FP8 Qwen3-VL-235B-A22B. The 30B-A3B MoE variant activates only 3B parameters yet is claimed to match or exceed GPT-5-Mini and Claude 4-Sonnet across STEM, VQA, OCR, video understanding, and agent tasks. Weights were released simultaneously on ModelScope, Hugging Face, and Qwen Chat, giving developers free access to the exact Qwen3-VL-30B-A3B-Instruct variant. The FP8 quantizations target more efficient deployment while preserving the compact MoE design. This release makes the small-footprint VLM broadly available for research and production use.

Alibaba's Qwen3-VL-30B-A3B-Instruct model card documents the MoE vision-language release with native 256K context expandable to 1M tokens. Architectural updates include Interleaved-MRoPE for long-horizon video reasoning, DeepStack ViT feature fusion, and Text–Timestamp Alignment for precise event localization in video. The card positions it as the strongest VLM in the Qwen series to date. Key capability upgrades listed in the model card include a Visual Agent that operates PC and mobile GUIs, visual code generation from images or videos, advanced 2D and 3D spatial grounding, and OCR expanded to 32 languages. Enhanced multimodal reasoning targets STEM and math, while text understanding is described as on par with pure LLMs, enabling lossless text–vision fusion for unified comprehension.

vLLM Ascend officially validates Qwen3-VL-30B-A3B-Instruct on Huawei Ascend hardware in the v0.13.0 cycle, treating it as a sparse MoE vision-language model with about 30B total parameters and roughly 3B activated per token. The tutorial covers supported features, image and video online deployment, offline inference, functional verification, and accuracy and performance evaluation, with deployment guidance for Atlas 800 A3 and A2 nodes. Both BF16 and w8a8-mxfp8 quantized weights are supported, giving Ascend users a ready-to-run multimodal stack. The documentation walks through Docker image and source installation, then explains shared multi-node weight deployment paths so that distributed serving jobs can locate the same model directory. It links into Ascend-specific feature configuration and performance tuning for long-context multimodal workloads. This positions Qwen3-VL-30B-A3B-Instruct as a first-class Ascend target alongside mainstream vLLM-validated models for vision-language serving.

Qwen3-VL-30B-A3B-Instruct is the official weight release of the Qwen3-VL generation, described on the model card as the most powerful vision-language model in the series to date, available in both Dense and MoE variants with Instruct and Thinking editions. Key upgrades include a Visual Agent that operates PC and mobile GUIs, visual coding that generates Draw.io, HTML, CSS, and JS from images or video, and stronger 2D and 3D spatial grounding for embodied tasks. It expands OCR to 32 languages from 19, handles rare scripts and jargon, and keeps text understanding on par with pure LLMs. Architecturally, the model introduces Interleaved-MRoPE over time, width, and height for long-horizon video reasoning, DeepStack fusion of multi-level ViT features for finer image-text alignment, and Text-Timestamp Alignment that replaces T-RoPE for precise event localization. It supports a native 256K context expandable to 1M, enabling hours-long video comprehension with second-level indexing, and excels at multimodal reasoning across STEM and math, outputting evidence-based conclusions for complex questions.

The Hugging Face model card for QuantTrio/Qwen3-VL-30B-A3B-Instruct-AWQ explicitly names the subject model Qwen/Qwen3-VL-30B-A3B-Instruct as the base, providing an AWQ quantization tailored for efficient deployment. The card documents a 17 GB model file (updated 2025-10-04) and supplies a vLLM startup command along wit On the architecture side, the card details Interleaved-MRoPE for full-frequency time/width/height positional allocation, DeepStack multi-level ViT feature fusion for fine-grained image-text alignment, and Text-Timestamp Alignment replacing T-RoPE for timestamp-grounded video reasoning. These technical specifics are tie