Qwen3-VL 30B A3B Instruct: A Vision Model You Can Host

The flagship of a model family gets the launch post. The mid-size member usually ends up doing the work. Qwen3-VL-30B-A3B-Instruct is that model for the Qwen3-VL line: small enough to run on hardware a normal team can own, cheap enough to point at a large document queue, and, on Qwen's own text numbers, not far behind the 235B on knowledge tasks.

Qwen released it on 4 October 2025, in Instruct and Thinking editions, under Apache-2.0, and published FP8 checkpoints for the Qwen3-VL models on the same day. This guide covers the Instruct edition, which answers directly without a reasoning trace.

The shape: 30B total, 3B active

It is a mixture-of-experts model. Hugging Face lists the checkpoint at 31B parameters including the vision encoder, and the name gives about 3B active per token. The technical report says the encoder follows the SigLIP-2 architecture, at the SO-400M size for every model except the two smallest. It shares the Qwen3-VL architecture with the 235B: interleaved MRoPE, DeepStack feature fusion and text timestamps for video, a 256K native context, and documented YaRN extension towards 1M.

The 3B active figure is the reason it is fast. The 31B total is the reason it still needs a real GPU: every expert has to be in memory even though few of them run on any given token.

What fits on what

By simple arithmetic, 31B parameters at 16-bit is about 62 GB of weights, and the FP8 checkpoint roughly halves that to around 31 GB. Add the KV cache, which grows with context and with the number of images per request, and the realistic single-GPU target is an 80 GB card: tight at 16-bit, comfortable with FP8. Qwen notes that its FP8 build needs H100-class hardware or newer, so older cards are limited to 16-bit or community quantisations.

Qwen also publishes GGUF builds for llama.cpp-based runtimes if you want to go smaller, with the usual quality cost; quantization explained covers where that cost tends to land. Qwen's serving examples cover vLLM 0.11.0 or later and SGLang, both with OpenAI-compatible APIs, so swapping it in behind existing client code is a configuration change.

That combination matters most when documents cannot leave your network. Medical scans, contracts and internal dashboards are the cases where a self-hosted 30B beats a hosted 235B on the only criterion that counts.

How much you give up against the 235B

The technical report publishes text-only scores for both Instruct models:

• MMLU-Pro: 77.8 for the 30B, 81.8 for the 235B
• AIME25: 69.3 against 74.7
• LiveCodeBench v6: 42.6 against 54.3

The knowledge gap is small and the coding gap is large. For the vision side, the report describes the 30B as delivering competitive results on multimodal reasoning without the headline claims it makes for the 235B. A fair reading: for reading text out of images and describing what is on screen, the gap will often not show; for reasoning about what is in the image and writing code from it, it will.

It also holds its own against the text-only model of the same size. Qwen's published figures for Qwen3-30B-A3B-Instruct-2507 are 78.4 on MMLU-Pro, 61.3 on AIME25 and 43.2 on LiveCodeBench v6, so adding vision did not cost it measurable text ability on these rows. Arena-Hard v2, a chat-preference test, is the exception in the same table: 58.5 against 69.0.

The price of a page

OpenRouter listed it at $0.15 input and $0.60 output per million tokens at the time of writing. For vision work the input side usually dominates, because Qwen's repository example budgets 256 to 1,280 visual tokens per image, scaled by resolution.

A worked example: a pipeline extracting fields from scanned invoices at the high end of that range, with a 500-token prompt and a 300-token JSON reply, comes to under 50 cents per thousand pages at those rates. The same job at the 235B Instruct's listed rates costs about 40% more on input and three times as much on output, roughly double overall. Small per page; not small across a large archive.

It supports JSON mode, structured outputs and tool calling, which is what makes extraction pipelines reliable. Structured outputs and JSON mode covers why constrained decoding beats asking nicely.

Picking between the Qwen3-VL editions

• This model for high-volume extraction, screenshot triage, captioning, private documents and anything you want on your own hardware.
• Qwen3-VL 30B Thinking for the subset of requests that need multi-step reasoning about the image, at the same hardware footprint.
• Qwen3-VL 235B Instruct for hard visual coding, dense documents where field-level accuracy failed on the 30B, and long video.

Recommended sampling for the Instruct editions is temperature 0.7, top-p 0.8, top-k 20 and a presence penalty of 1.5. Start there, then lower the temperature for extraction, where you want the same answer every time. Measure field-level accuracy on fifty of your own documents before deciding the 30B is not enough.