Language modelActive

Qwen 3 VL 235B A22B Thinking

Alibaba

Released: September 22, 2025
Data date: September 8, 2026

Speed: 55.8 tokens/s
Output throughput
Price: $0.4 · $4
Input · Output / 1M tokens
Input: Text, Image, Video
Output: Text

Qwen 3 VL 235B A22B Thinking is the large reasoning checkpoint in the Qwen3-VL family. It combines text, image, and video understanding with a trained thinking mode. Its Mixture-of-Experts architecture contains 235 billion parameters but activates 22 billion per token. The technical report gives it a 256K context for interleaved text, image, and video inputs.

Qwen3 VL 235B A22B Instruct is a separate checkpoint for direct answers without this Thinking variant’s reasoning trace.

Specifications and access

Specification	Value and source
Weights	Open weightsSource
Access	API, LocalSource
Parameters	235BSource
Active parameters	22BSource
Context window	262,144 tokensSource
Weights repository	Qwen/Qwen3-VL-235B-A22B-ThinkingSource
License	Apache-2.0Source
Architecture	vision encoder + MLP vision-language merger + Qwen3 language model; dense or MoESource
Input	Text, Image, VideoSource
Output	TextSource
Reasoning mode	thinkingSource
Tool calling	YesSource
Structured output	YesSource
Native context window	262,144 tokensSource
Maximum context window	262,144 tokensSource
Parameters	235 BSource
Active parameters	22 BSource
Languages	39 languages testedSource
Repository created	2025-09-22 (Repository creation date is a precise publication proxy, but may differ from the public announcement date or an individual checkpoint's first availability.)Source

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

MMMU (Show measurement, test conditions, and source)

Source value: 80.6
Score: 80.6
Metric: percent
Unit: %
Category: visual
Direction: Higher is better
Source type: vendor-reported
Evaluator: Qwen3-VL Technical Report
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Exact model column; report evaluates thinking/instruct variants separately

Source Qwen3-VL Technical Report

MMMU-Pro (Show measurement, test conditions, and source)

Source value: 69.3
Score: 69.3
Metric: percent
Unit: %
Category: visual
Direction: Higher is better
Source type: vendor-reported
Evaluator: Qwen3-VL Technical Report
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Exact model column; report evaluates thinking/instruct variants separately

Source Qwen3-VL Technical Report

MathVision (Show measurement, test conditions, and source)

Source value: 74.6
Score: 74.6
Metric: percent
Unit: %
Category: visual
Direction: Higher is better
Source type: vendor-reported
Evaluator: Qwen3-VL Technical Report
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Exact model column; report evaluates thinking/instruct variants separately

Source Qwen3-VL Technical Report

DocVQA test (Show measurement, test conditions, and source)

Source value: 96.5
Score: 96.5
Metric: percent
Unit: %
Category: document
Direction: Higher is better
Source type: vendor-reported
Evaluator: Qwen3-VL Technical Report
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Exact model column; report evaluates thinking/instruct variants separately

Source Qwen3-VL Technical Report

VideoMME w/o subtitles (Show measurement, test conditions, and source)

Source value: 79
Score: 79
Metric: percent
Unit: %
Category: video
Direction: Higher is better
Source type: vendor-reported
Evaluator: Qwen3-VL Technical Report
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Exact model column; report evaluates thinking/instruct variants separately

Source Qwen3-VL Technical Report

OCRBench (Show measurement, test conditions, and source)

Source value: 875
Score: 875
Metric: points
Unit: points
Category: document
Direction: Higher is better
Source type: vendor-reported
Evaluator: Qwen3-VL Technical Report
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Exact model column; report evaluates thinking/instruct variants separately

Source Qwen3-VL Technical Report

Output speed (Show measurement, test conditions, and source)

Source value: 55.8351013442385
Score: 55.8351013442385
Metric: tokens/s
Unit: tokens/s
Direction: Higher is better
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-09-08
Methodology: Independent evaluation. Reasoning; default 10K-input workload only
Context: Artificial Analysis performance timing; output speed is decoded after the first streaming chunk. Provider basis: first-party API. Deprecated model. Artificial Analysis only continues benchmarking the default 10K-input workload.

Source Artificial Analysis

Latency reasoning time (Show measurement, test conditions, and source)

Source value: 35.819761258593566
Score: 35.819761258593566
Metric: seconds
Unit: seconds
Direction: Lower is better
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-09-08
Methodology: Independent evaluation. Reasoning; default 10K-input workload only
Context: Artificial Analysis performance timing; output speed is decoded after the first streaming chunk. Provider basis: first-party API. Deprecated model. Artificial Analysis only continues benchmarking the default 10K-input workload.

Source Artificial Analysis

Latency input time (Show measurement, test conditions, and source)

Source value: 2.98233183900004
Score: 2.98233183900004
Metric: seconds
Unit: seconds
Direction: Lower is better
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-09-08
Methodology: Independent evaluation. Reasoning; default 10K-input workload only
Context: Artificial Analysis performance timing; output speed is decoded after the first streaming chunk. Provider basis: first-party API. Deprecated model. Artificial Analysis only continues benchmarking the default 10K-input workload.

Source Artificial Analysis

Answer generation time (end-to-end timing breakdown) (Show measurement, test conditions, and source)

Source value: 8.954940314648391
Score: 8.954940314648391
Metric: seconds
Unit: seconds
Direction: Lower is better
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-09-08
Methodology: Independent evaluation. Reasoning; default 10K-input workload only
Context: Artificial Analysis performance timing; output speed is decoded after the first streaming chunk. Provider basis: first-party API. Deprecated model. Artificial Analysis only continues benchmarking the default 10K-input workload.

Source Artificial Analysis

End-to-end reasoning time component (Show measurement, test conditions, and source)

Source value: 35.819761258593566
Score: 35.819761258593566
Metric: seconds
Unit: seconds
Direction: Lower is better
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-09-08
Methodology: Independent evaluation. Reasoning; default 10K-input workload only
Context: Artificial Analysis performance timing; output speed is decoded after the first streaming chunk. Provider basis: first-party API. Deprecated model. Artificial Analysis only continues benchmarking the default 10K-input workload.

Source Artificial Analysis

End-to-end input time component (Show measurement, test conditions, and source)

Source value: 2.98233183900004
Score: 2.98233183900004
Metric: seconds
Unit: seconds
Direction: Lower is better
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-09-08
Methodology: Independent evaluation. Reasoning; default 10K-input workload only
Context: Artificial Analysis performance timing; output speed is decoded after the first streaming chunk. Provider basis: first-party API. Deprecated model. Artificial Analysis only continues benchmarking the default 10K-input workload.

Source Artificial Analysis

Sources and data date

Type	Evidence and data date
Speed	Alibaba, Reasoning. Output speed after the first streaming chunk, not total response time. Classified as deprecated by Artificial Analysis. Artificial Analysis maintains only the default 10K-input workload there. Retrieved September 8, 2026. Source: Artificial Analysis · Artificial Analysis (retrieved September 8, 2026)Speed measurement methodology
Price	Base API prices without caching or batch discounts. Higher context tiers and other rates are listed under Costs. Source: Pricing · API pricing
Input	Documented modalities. Retrieved September 8, 2026. Source: Model documentation · Model metadata (retrieved September 8, 2026) · Qwen3-VL-235B-A22B-Thinking model card · Editorial description reviewed October 2, 2026
Output	Documented modalities. Retrieved September 8, 2026. Source: Model documentation · Model metadata (retrieved September 8, 2026) · Qwen3-VL-235B-A22B-Thinking model card · Editorial description reviewed October 2, 2026
Research status	Research date September 8, 2026. 4 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Additional source	Qwen3-VL Technical Report (retrieved September 8, 2026) · Qwen3-VL Technical Report · Editorial description reviewed October 2, 2026
Additional source	Model metadata (retrieved September 8, 2026)
Additional source	OpenRouter: Popularity index (derived), CC BY 4.0
Additional source	Vercel AI Gateway: Popularity index (derived), CC BY 4.0
Additional source	Creative Commons Attribution 4.0

Qwen 3 VL 235B A22B Thinking

Specifications and access

Measurements without matching peer values

Our tests

Pelican on a bicycle

Popularity

Published prices

Sources and data date