Qwen3 VL 235B A22B Instruct
Qwen · released Sep 22, 2025 · Qwen/Qwen3-VL-235B-A22B-Instruct
- Type
- Open weightsApache License 2.0
- Params
- 236B
- Context
- 262K
22B active per word · about 197K words of context
Our take
Written Sep 17, 2026Qwen3 VL 235B A22B Instruct takes text and images in and returns text, and you can download it and run it yourself. Only about 22 billion of its 235.7 billion parameters work on any one token, so the memory it needs is far below what the total size suggests. The only quality evidence we list is human preference, not correctness.
Use it for work that mixes pictures and text — reading screenshots, diagrams or scanned pages alongside a written question — or when you want to run a vision model yourself on hardware sized for its active footprint. The licence allows commercial use, changes and redistribution (Apache 2.0). Skip it if you need a measured coding or reasoning score rather than a preference rating, or if you need a small model that fits a single consumer card.
The case for it
- About 22 billion of its 235.7 billion parameters work per token, so the memory in use is closer to a mid-size model than to a 235-billion-parameter one — running it yourself is plausible.
- The licence allows commercial use, changes and redistribution (Apache 2.0).
- Text and images go into the same request, so a screenshot or diagram does not have to be described in words first.
- The request capacity leaves room for long documents and images together without splitting them up first, though reliable recall across all of it is unverified in our data.
The case against it
- The only quality evidence we list is human pairwise preference across six arena categories, which records which answer people preferred rather than whether it was correct.
- No serving speed is supplied for any of the nine hosted offers, so price alone cannot pick the host.
- The cheapest input rate and the cheapest output rate are not the same offer: DeepInfra is cheapest on both, while Parasail, Venice and OpenRouter charge more than double DeepInfra's output rate.
How good is it?
EverydayGeneral questions and everyday reasoning
Arena Text (overall)87th of 168 · 1413
CodingWriting and fixing code on its own
Arena Coding81st of 168 · 1463
Arena Coding is the only board that has scored it for this.
AgenticPlanning, calling tools, staying on task
Not yet scored on Arena Agent.
WritingDrafting and rewriting prose
Arena Creative Writing97th of 168 · 1357
Arena Creative Writing is the only board that has scored it for this.
These tests check whether a model follows instructions — a precondition for all the work above, but not a measure of how well that work is done.
Every published score for this model6 scoresEvery figure we hold, from 6 boards, with who ran it and a link to the source — including the boards no rating above is built on.
Can you run it yourself?
GeForce RTX 4090 · 24 GB
Too large for this card. The weights do not fit even with part of them offloaded to system memory.
Apple M1 Pro (16-core GPU) · 32 GB
Too large for this card. The weights do not fit even with part of them offloaded to system memory.
Comfortable fit
Apple M3 Ultra (80-core GPU) · 512 GB
Room to spare. 229.3 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 4 hours ago — each listing carries its own date.
The only listing at 262K of context — the other 5 in the table below are not like-for-like. 3 cheaper rows there are outside that comparison: a different context length or a different quantisation.
- per 1M tokens
- $0.21 in / $1.90 out
- Context served
- 262K
- Throughput
- Not measured
| Provider | In / out per 1M tokens | Context | Throughput | Trains on prompts | Logs prompts | Zero retention |
|---|---|---|---|---|---|---|
| DeepInfrafp8Direct and through OpenRouter | $0.20 / $0.88checked 4 hours ago | 262K16K max reply through OpenRouter | 8 tok/sthrough OpenRouter | DirectUnknownThrough OpenRouterNo | DirectUnknownThrough OpenRouterNo | DirectUnknownThrough OpenRouterConfirmed |
| Alibaba CloudThrough OpenRouter | $0.26 / $1.04checked 4 hours ago | 131K33K max reply | 39 tok/s | No | Yesunknown period | Unknown |
| Novita AIbf16Direct and through OpenRouter | $0.30 / $1.50checked 4 hours ago | 131K33K max reply through OpenRouter | 45 tok/sthrough OpenRouter | DirectUnknownThrough OpenRouterNo | DirectUnknownThrough OpenRouterNo | DirectUnknownThrough OpenRouterConfirmed |
| Parasailfp8Through OpenRouter | $0.21 / $1.90checked 4 hours ago | 131K33K max reply | 22 tok/s | No | No | Confirmed |
| Venice AIfp8Through OpenRouter | $0.21 / $1.90checked 4 hours ago | 128K16K max reply | 33 tok/s | No | No | Confirmed |
| OpenRouterOpenRouter's own listing | $0.21 / $1.90checked 4 hours ago | 262K | not measured | Unknown | Unknown | Unknown |
DeepInfra
Novita AIAcross the 6 listings we hold: 5 say they do not train on prompts (2 of them only through OpenRouter), 0 say they do and 1 does not say. 4 appear in the zero-retention registry we check (2 of them only through OpenRouter); 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 |
|---|---|---|---|
| DeepInfrafp8Direct and through OpenRouter | ✓ | ✓ | ✓ |
| Alibaba CloudThrough OpenRouter | ✓ | ✓ | ✓ |
| Novita AIbf16Direct and through OpenRouter | ✓ | ✓ | ✓ |
| Parasailfp8Through OpenRouter | ✓ | ✓ | ✓ |
| Venice AIfp8Through OpenRouter | ✓ | ✓ | ✓ |
| OpenRouterOpenRouter's own listing | ✓ | ✓ | ✓ |
Tool calling: 6 of 6 listings say yes. JSON output: 6 of 6 listings say yes. Strict schema: 6 of 6 listings say yes.
Models people weigh against Qwen3 VL 235B A22B Instruct
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.
- Nothing we hold says whether an endpoint streams, so we do not show it either way.
- 1 of 6 listings does not say whether it trains on prompts, and 2 answer only through OpenRouter, not for their own listing.
- 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-235B-A22B-Instruct
- Architecture
- Mixture of experts
- Takes in, gives back
- Text and images in, text out
- Catalogue slug
- qwen-qwen3-vl-235b-a22b-instruct