Qwen3 VL 235B
Alibaba's 235B-parameter open-weights vision-language MoE with tool use, web search, and visual agent capabilities.
Overview
What is Qwen3 VL 235B
Qwen3 VL 235B is Alibaba Cloud's open-weights vision-language model built on a Mixture-of-Experts architecture. It processes text, images, and video, supports tool use and web search, and offers strong spatial reasoning and visual coding capabilities. The 235B-parameter MoE runs with 22B active parameters and is licensed under Apache 2.0.
Running it privately on Venice
On Venice, Qwen3 VL 235B runs under a private, zero-retention privacy tier — your prompts and images are not stored, profiled, or used for training. You get native vision, tool use, web search, and structured JSON output on an open-weights Apache 2.0 model, with no Big-Tech surveillance or account-linked history.
Assessment
Strengths and limitations
- Open-weight Apache 2.0 vision-language model with native tool use, web search, structured JSON output, and multiple-image support.
- Strong visual reasoning and spatial perception, including GUI agent operation, visual coding from screenshots, and advanced OCR across 32 languages.
- Efficient MoE inference: 235B total parameters with 22B active per forward pass, delivering high-end multimodal quality without dense-model compute costs.
- Handles interleaved text-image-video contexts for complex document analysis and video indexing workflows.
- The Instruct variant prioritizes speed and cost over maximum reasoning depth; for complex STEM or causal analysis the Thinking edition may perform better.
- On Venice it is quantized to fp8 and served at a 128K context window, which may limit very long-document or hour-long video tasks compared to the native 256K–1M range cited in upstream docs.
- While open weights, it is not uncensored: standard safety alignment is still present.
- Vision-language latency is higher than text-only models, and output costs are nearly $2 per million tokens, so high-volume text-only workloads may be cheaper on DeepSeek V3.2 or Gemma 4.
Capabilities
What it supports
- Tool use / function calling
- Vision (image input)
- Reasoning
- Web search
- Code-optimized
- Structured output (JSON schema)
- Audio input
- Video input
- Multiple image inputs
- Log probabilities
Specifications
Datasheet
- Maker
- Alibaba Cloud (Qwen team)
- Released
- September 2025
- Architecture
- Mixture of Experts (MoE)
- Parameters
- 235B total, 22B active
- Open weights
- Yes — Apache 2.0
- Context window
- 128K tokens
- Max output
- 16.384K tokens
- Capabilities
- Vision, Function calling, Web search
- Privacy on Venice
- Private — zero retention
- Available on Venice since
- Jan 2026
API
Call it from your code
Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen3-vl-235b-a22b",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Pricing
What it costs on Venice
Billed per token on Venice: $0.21 per 1M input tokens and $1.90 per 1M output tokens.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
Alternatives
How it compares
| Model | Context window | Open weights | Price (Venice) | Best for |
|---|---|---|---|---|
| Qwen3 VL 235B | 128K tokens | Yes | $0.21 in · $1.90 out / 1M | Vision, tool use & web search |
| DeepSeek V3.2 | 160K tokens | Yes | $0.33 in · $0.48 out / 1M | Text reasoning & coding |
| Kimi K2.6 | 256K tokens | Yes | $0.75 in · $3.50 out / 1M | Long-context text |
| Claude Sonnet 4.6 | 1M tokens | No | $3.60 in · $18 out / 1M | General multimodal |
The only open-weights vision model in this set with native tool use, web search, and multiple-image support at under $2/1M output.
Use cases
What it is good for
- 01Visual document analysis and OCR for invoices, forms, and long scans in 32 languages.
- 02GUI automation and visual agent tasks on desktop or mobile interfaces.
- 03Front-end coding from design mockups: generating HTML, CSS, or JS from screenshots.
- 04Multimodal data extraction and structured cataloging from product images or video frames.
- 05Grounded research with web search over current visual content and news.
Prompting
Getting better results
Use interleaved image-text prompts for document analysis; place the image first, then ask specific questions about regions or text.
For visual coding, explicitly state the desired output format (e.g., 'Generate HTML/CSS from this screenshot').
Leverage tool use and web search when you need real-time facts to interpret images or verify visual claims.
Upload multiple images in one conversation to compare products, track changes across frames, or assemble a visual narrative.
Version history
Standard instruction-tuned variant with vision, tool use, and web search.
Reasoning-enhanced edition for deeper multimodal reasoning and STEM tasks.
FAQ
Frequently asked questions
Qwen3 VL 235B is Alibaba Cloud's open-weights vision-language model released in September 2025. Built on a Mixture-of-Experts architecture with 235B total and 22B active parameters, it understands text, images, and video, supports tool use and web search, and excels at visual reasoning, OCR, and GUI automation.
Venice charges $0.21 per 1M input tokens and $1.90 per 1M output tokens. Cached input is $0.10 per 1M. There is no subscription; you pay only for the tokens you use.
Yes. The weights are released under the Apache 2.0 license by the Qwen team, permitting commercial use and self-hosting. On Venice it runs as open weights with fp8 quantization.
New Venice accounts include free credits to try the model, with no credit card required. Beyond the free tier, usage is billed per token at the published rates.
Yes. On Venice it supports function calling, structured JSON output, web search, and multiple image inputs, making it suitable for agent workflows and automated visual tasks.
Choose Qwen3 VL 235B for image, video, and multimodal document tasks; it offers vision and visual agent capabilities that DeepSeek V3.2 does not. Choose DeepSeek V3.2 for pure text reasoning and coding at lower cost if you do not need vision.
It can describe, analyze, and reason over images; perform OCR in 32 languages; generate code from UI screenshots; judge spatial relationships and occlusions; and process multiple images in a single conversation.
Yes. On Venice the model has web search capability, so it can ground answers in current information beyond its training cutoff.
Run Qwen3 VL 235B privately
No prompt logging. No data used for training.