Qwen3 VL 30B A3B Thinking
Qwen · released Sep 30, 2025 · Qwen/Qwen3-VL-30B-A3B-Thinking
- Type
- Open weightsApache License 2.0
- Params
- 31.1B
- Context
- 262K
3B active per word · about 197K words of context
Our take
Written Sep 6, 2026Qwen3 VL is a vision-language model that can reason through problems before answering, with only 3 billion active parameters drawn from 31.1 billion total. Its Apache licence and low active compute make it attractive for edge deployment, though no benchmark scores verify its quality yet.
Pick this for local or edge vision-language tasks where active compute must stay minimal, or for budget hosted inference with a permissive licence. Use it when you need redistribution rights or fine-tuning freedom. Skip it if you need verified quality scores, consistent throughput across providers, or predictable pricing — the same model costs markedly more at some hosts.
The case for it
- Extreme parameter efficiency: 3 billion active from 31.1 billion total, roughly a 10:1 sparsity ratio.
- Apache 2.0 licence allows commercial use, fine-tuning and redistribution.
- 262,144-token request limit is large for its active-parameter class.
- Output price varies widely by host — the cheapest is well under half the rate of the most expensive.
The case against it
- No benchmark scores recorded: chat, reasoning, vision and coding performance are all unverified.
- Throughput data is sparse and inconsistent, with most providers not disclosing speed.
- Identical model, identical input rate, yet some major providers charge more than double the output price of others.
How good is it?
We hold no score for this model.
So there is no figure here for everyday use, coding, agent work or writing. That is a gap in our data, not a low score.
Can you run it yourself?
GeForce RTX 4090 · 24 GB
Borderline fit on an estimated size. It leaves 1.2 GB spare on a size we calculated rather than measured, and a 10% error either way would change the answer.
Comfortable fit
GeForce RTX 5090 · 32 GB
Room to spare. 9.2 GB spare means a 10% error in the size would not change the answer.
Apple M1 Pro (16-core GPU) · 32 GB
Room to spare. 2.4 GB spare means a 10% error in the size would not change the answer.
Memory use by level
Against a 24 GB card.
What is quantisation? →This model on every device we track71 devicesThe Q4 build most people download, on each device: what the weights come to, how much context the memory leaves, and whether it runs. Smallest device that runs it first.
Check against your own machine → · Where to rent it hosted →
Or rent it from someone else
Prices checked between 2 hours and 21 days ago — each listing carries its own date.
The only listing at 262K of context — the other 3 in the table below are not like-for-like. 2 cheaper rows there are outside that comparison: a different context length or a different quantisation.
- per 1M tokens
- $0.20 in / $2.40 out
- Context served
- 262K
- Throughput
- Not measured
| Provider | In / out per 1M tokens | Context | Throughput | Trains on prompts | Logs prompts | Zero retention |
|---|---|---|---|---|---|---|
| Novita AIDirect | $0.20 / $1.00checked 21 days ago | 131K | not measured | Unknown | Unknown | Unknown |
| SiliconFlowfp8Through OpenRouter | $0.29 / $1.00checked 2 hours ago | 262K236K max reply | 83 tok/s | No | No | Confirmed |
| OpenRouterOpenRouter's own listing | $0.20 / $2.40checked 2 hours ago | 262K | not measured | Unknown | Unknown | Unknown |
| Alibaba CloudThrough OpenRouter | $0.20 / $2.40checked 2 hours ago | 131K33K max reply | 113 tok/s | No | Yesunknown period | Unknown |
Novita AIAcross the 4 listings we hold: 2 say they do not train on prompts, 0 say they do and 2 do not say. 1 appears in the zero-retention registry we check; the rest are unknown to us.
What each host's API supports
From the parameter list each endpoint publishes. Streaming is omitted: nothing we hold reports it, for any model.
| Provider | Tool calling | JSON output | Strict schema |
|---|---|---|---|
| Novita AIDirect | |||
| SiliconFlowfp8Through OpenRouter | ✓ | ✓ | ✓ |
| OpenRouterOpenRouter's own listing | ✓ | ✓ | ✓ |
| Alibaba CloudThrough OpenRouter | ✓ | ✓ | ✓ |
Tool calling: 3 of 4 listings say yes, 1 publishes no parameter list. JSON output: 3 of 4 listings say yes, 1 publishes no parameter list. Strict schema: 3 of 4 listings say yes, 1 publishes no parameter list.
When we formed this view
Recent changes
What moved
first indexed by our pipelineEach date is the day we first saw the change, or the day the maker announced it.
What we do not know about this model yet
- We hold no measured file for it, so all 3 sizes on this page are calculated from the parameter count.
- No independent board has scored it, so we hold no quality figures at all.
- 1 of 4 listings publishes no parameter list, so what its API accepts is unknown to us.
- Nothing we hold says whether an endpoint streams, so we do not show it either way.
- 2 of 4 listings do not say whether they train on prompts.
- We hold no cached-input rate for any of its listings.
- We hold no batch or off-peak rate for any of its listings.
- We hold a decode speed for it, but no prompt-processing (prefill) figure, so how long the input side of a job takes is unknown to us.
Licence and identifiers
What the licence allowsApache License 2.0, what it allows commercially, and the identifiers you need to pull this model — its Hugging Face repo, our slug and a machine-readable card.
Licence
Apache License 2.0
Fully permissive: commercial use, redistribution, and derivatives allowed. Requires attribution and a copy of the license. Includes an express patent grant.
Identifiers
- Hugging Face
- Qwen/Qwen3-VL-30B-A3B-Thinking
- Architecture
- Mixture of experts
- Takes in, gives back
- Text and images in, text out
- Catalogue slug
- qwen-qwen3-vl-30b-a3b-thinking