Qwen: Qwen3 VL 235B A22B Instruct
Qwen3-VL-235B-A22B Instruct is an open-weight multimodal model that unifies strong text generation with visual understanding across images and video. The Instruct model targets general vision-language use (VQA, document parsing, chart/table...
Anyone in the Space can @-mention Qwen: Qwen3 VL 235B A22B Instruct with the team's shared context - pooled credits, one chat, one memory.
Starter is free forever - 1 Space, 100 credits/month, 1 MCP. No card.
Verdict
Best for
- Long-context document analysis with images
- Cost-sensitive multimodal applications
- Visual question answering on technical diagrams
- Screenshot analysis and UI understanding
- Batch processing of image-text pairs
Strengths
The 262K context window puts this model ahead of many vision-language competitors for processing lengthy documents with embedded images or analyzing multi-page PDFs. Pricing sits below premium models like GPT-4V while maintaining the ability to handle both modalities. The A22B active parameter count suggests efficient inference relative to the full 235B parameter base, potentially delivering faster responses than fully-dense models at this scale.
Trade-offs
Lack of public benchmarks means you're flying blind on comparative performance against Claude 3.5 Sonnet or GPT-4V on standard vision tasks. Proprietary licensing limits deployment flexibility compared to open-weight alternatives. The model is new enough that community knowledge around prompt engineering and failure modes hasn't accumulated. You'll need to run your own evals before trusting it in production, especially for safety-critical vision tasks.
Specifications
- Provider
- qwen
- Category
- llm
- Context length
- 262,144 tokens
- Max output
- 16,384 tokens
- Modalities
- text, image
- License
- proprietary
- Released
- 2025-09-23
Pricing
- Input
- $0.20/Mtok
- Output
- $0.88/Mtok
- Model ID
qwen/qwen3-vl-235b-a22b-instruct
Per-token prices show what the model costs upstream. On Switchy your team draws from one shared org credit pool - one plan, one balance for everyone.
Team cost calculator
5 seats · 80 msgs/day
Switchy meters this against your org's shared credit pool - one plan, one balance for everyone.
Providers
| Provider | Context | Input | Output | P50 latency | Throughput | 30d uptime |
|---|---|---|---|---|---|---|
| qwen | 262k | $0.20/Mtok | $0.88/Mtok | — | — | — |
Performance
Benchmarks
Works well with
Top MCPs
Compatibility data comes from first-party telemetry; once we have enough co-usage signal, top MCPs for this model will appear here.
How Switchy teams use it
Starter prompts
Technical Diagram Extraction
Analyze this technical diagram and extract all labeled components, their connections, and spatial relationships. Format your response as a structured list with component IDs, types, and connection paths.Open in a Space →
Multi-Page PDF Summarization
This PDF contains multiple pages of text, charts, and images. Provide a 200-word executive summary that integrates insights from all visual and textual elements, highlighting key data points and conclusions.Open in a Space →
UI Screenshot Analysis
Examine this application screenshot. List all interactive elements, identify any usability issues or accessibility concerns, and suggest three specific improvements to the layout or information hierarchy.Open in a Space →
Visual Data Extraction
Extract all data points from this chart or graph and output them as a CSV table. Include axis labels, legend entries, and any annotations visible in the image.Open in a Space →
Document Classification with Images
Classify this document based on both its text content and any images present. Assign it to one of these categories: [list your categories]. Explain which visual and textual features drove your classification.Open in a Space →
Example outputs
Illustrative - representative of the model's voice and quality, not literal recordings.
Analyze this architectural blueprint image and identify potential structural issues or code violations. Focus on load-bearing elements and egress requirements.
In this illustrative example, the model would examine the blueprint and produce a detailed structural assessment. It would identify that the second-floor joists span 18 feet without intermediate support, exceeding typical residential code limits of 16 feet for 2x10 lumber at 16-inch centers. The model would note that the stairwell width measures only 34 inches, falling short of the 36-inch minimum required by most building codes. It would also flag that the load-bearing wall removal on the ground floor lacks a specified beam size, recommending consultation with a structural engineer for proper header dimensioning based on the 22-foot span and expected loads.
With 262k token context, this model can process high-resolution architectural drawings alongside lengthy building codes. The multimodal capability lets it cross-reference visual elements with regulatory text. At $0.88/Mtok output, extended technical reports become expensive quickly — best suited for high-value professional reviews rather than casual queries.
I have 47 customer support tickets from last week. Extract common themes, sentiment trends, and flag any urgent technical issues that need escalation. Here's the full transcript dump...
In this illustrative scenario, the model would process all 47 tickets within its context window and produce a structured analysis. It would identify three dominant themes: login authentication failures (19 tickets, mostly frustrated tone), billing discrepancy questions (15 tickets, neutral-to-negative sentiment), and feature requests for mobile export (13 tickets, positive engagement). The model would flag two urgent cases: a data sync failure affecting a enterprise client's production environment, and a security concern where a user reported unauthorized access attempts. It would provide ticket IDs, severity scores, and recommended next actions for each escalation.
The 262k context window handles batch analysis that would require multiple API calls with smaller models. This reduces latency and maintains coherence across the full dataset. However, at $0.20/Mtok input, processing 47 full transcripts regularly adds up — consider summarizing tickets first for routine analysis, reserving full-context runs for complex investigations.
Write a Python function that takes a pandas DataFrame of time-series sensor data and detects anomalies using isolation forest. Include proper error handling and docstrings.
In this illustrative example, the model would generate a complete, production-ready function with comprehensive error handling. The code would validate input DataFrame structure, check for required columns, handle missing values appropriately, and raise informative exceptions for edge cases. The isolation forest implementation would include configurable contamination parameters and return both anomaly labels and anomaly scores. The docstring would follow NumPy style, documenting parameters with type hints, return values, potential exceptions, and a usage example. The model would also include inline comments explaining the isolation forest's contamination threshold choice and why certain preprocessing steps matter for sensor data specifically.
The model's text-generation capability produces well-structured code with attention to real-world engineering concerns like error handling and documentation. The 235B parameter count suggests strong reasoning about code patterns and best practices. Without public benchmarks, it's unclear how this compares to specialized code models on complex refactoring or bug-finding tasks — evaluation on your specific codebase is recommended.
Use-case deep-dives
When 262K context beats OCR pipelines for document teams
A 4-person accounting ops team processing 200+ vendor invoices daily hits the sweet spot here. Qwen3 VL 235B handles the full invoice stack—line items, tables, handwritten notes, multi-page PDFs—in one 262K-token context window at $0.20/Mtok input. That's roughly $0.05 per 100-page batch, which undercuts most OCR-plus-LLM chains once you factor in preprocessing costs. The vision layer reads scanned tables directly; no Tesseract step. Output at $0.88/Mtok means structured JSON extraction runs $0.15-0.25 per invoice depending on schema complexity. If your invoices average under 50 pages and you're processing fewer than 100/day, a smaller vision model saves money. Above that threshold, the context window and input pricing make this the default choice for document-heavy workflows.
Why this model scales for high-volume image moderation
A 12-person marketplace team reviewing 5,000 product photos daily needs speed and cost predictability. Qwen3 VL 235B processes image-plus-text prompts at $0.20/Mtok input, which translates to roughly $0.002-0.004 per image when you're checking compliance (background color, logo placement, prohibited items). The 262K context lets you batch 50-100 images in a single call with a shared rubric, cutting API overhead by 98% versus per-image requests. Output tokens are minimal for yes/no decisions, so the $0.88/Mtok rate stays under $0.01 per batch. No public benchmarks yet, so run a 500-image pilot before committing. If accuracy on edge cases (reflective surfaces, low-light shots) falls below 92%, keep a human-in-loop for flagged items. For clean product photography at this volume, the economics work.
When vision context replaces multi-turn debugging threads
A 20-person SaaS support team handling 800 tickets/day with screenshots attached saves 6-8 minutes per ticket by front-loading diagnosis. Qwen3 VL 235B reads the screenshot, error message, and 10-15 prior chat turns (all in one 262K context) to route the ticket and draft a first-response suggestion. Input cost is $0.20/Mtok, so a typical ticket with 2 screenshots and 3K tokens of history runs $0.0008. Output at $0.88/Mtok adds $0.003-0.005 for a 400-word draft. That's under $0.006 per ticket, or $4.80/day at 800 tickets. The model hasn't published accuracy benchmarks, so measure false-route rate in week one. If it's above 12%, restrict to low-complexity categories (password resets, UI navigation). Below that, expand to billing and integration issues.
Frequently asked
Is Qwen3 VL 235B good for vision-language tasks?
Yes, Qwen3 VL 235B is built for multimodal work — text and image inputs together. The 235B parameter count suggests strong reasoning capability across both modalities. With 262k context window, you can feed multiple high-res images plus long text prompts in one request. No public benchmarks yet, so you're relying on Qwen's track record from earlier releases.
Is Qwen3 VL 235B cheaper than GPT-4 Vision?
Yes, significantly. At $0.20 input and $0.88 output per million tokens, Qwen3 VL costs roughly 85% less than GPT-4 Vision for most workloads. The trade-off is maturity — GPT-4V has years of production hardening and published benchmarks. If cost matters more than brand safety, Qwen3 VL wins on price alone.
Can Qwen3 VL 235B handle 100+ page PDFs with images?
Probably. The 262k context window theoretically fits 100+ pages of mixed text and images, depending on image resolution and token encoding. Real-world performance depends on how the model handles long-context retrieval — no published evals yet. Test with your actual documents before committing to production; context window size doesn't guarantee quality at the edges.
How does Qwen3 VL 235B compare to Qwen2.5 VL?
Qwen3 VL 235B is the successor, likely with better reasoning and longer context than Qwen2.5 VL's 128k window. The parameter jump from Qwen2.5's 72B to 235B suggests major capability gains, especially for complex visual reasoning. Without head-to-head benchmarks, assume incremental improvements in accuracy and instruction-following, not a generational leap.
Should I use Qwen3 VL 235B for real-time image analysis?
No, not for sub-second latency needs. At 235B parameters, inference will take multiple seconds per request even on optimised infrastructure. Use this for batch document processing, detailed image captioning, or multimodal RAG where 2-5 second response times are acceptable. For real-time vision, drop to smaller models like Qwen2-VL-7B or vision-specific APIs.