Qwen3 VL 8B

qwenqwen/qwen3-vl-8b

Qwen3 VL 8B — compact vision-language model for image understanding tasks.

Hugging Face model card (Instruct) ↗Qwen3-VL GitHub repository ↗Qwen3-VL technical report (arXiv 2511.21631) ↗Qwen Chat demo ↗

Qwen3-VL-8B is Alibaba Qwen's compact dense vision-language model in the Qwen3-VL family, released on October 15, 2025. It shares the same architectural DNA as its larger siblings — Interleaved-MRoPE for temporal-spatial positional encoding across interleaved image/video/text sequences, DeepStack for multi-level ViT feature fusion, and Text–Timestamp Alignment for precise video temporal grounding — packed into an 8B-parameter form factor suited for cost-sensitive production deployments, edge inference, and latency-critical pipelines. Available in two editions: Instruct (instruction-tuned, standard completion) and Thinking (reasoning-enhanced, with step-by-step CoT traces), both accessible on TheRouter through the model ID qwen/qwen3-vl-8b.

For TheRouter users, Qwen3-VL-8B is the go-to vision-language model when per-token cost or deployment footprint matters more than peak reasoning depth. It achieves strong real-world results — DocVQA 96.1%, ScreenSpot 94.4%, OCRBench 89.6% — that match or surpass many larger models, making it practical for high-volume document processing, GUI automation, and OCR workloads across 32 languages. The Thinking edition adds chain-of-thought reasoning, lifting MMMU and MathVista scores substantially for math-embedded and multi-step visual tasks where a larger model is otherwise needed.

Best for

• High-volume document OCR and extraction — 32-language OCR with robustness to low light, blur, and tilt; structured data extraction from invoices, forms, and receipts at scale where cost per page is a primary constraint
• GUI and visual agent automation — identifies UI elements, understands their functions, invokes tools, and completes tasks on PC/mobile screens; ScreenSpot 94.4% makes it competitive for screenshot-driven agentic pipelines without 32B cost
• Cost-sensitive image QA and classification — product recognition, celebrity/landmark/flora-fauna identification, chart and table comprehension for applications where throughput and margin matter over marginal accuracy gains
• Thinking edition for compact STEM reasoning — use when you need math-embedded visual problem solving (MathVista lift with CoT) but cannot afford 32B or 235B inference cost; ideal for edtech and scientific image annotation pipelines
• Edge and latency-constrained deployments — smaller memory footprint than 32B makes Qwen3-VL-8B deployable on single-GPU or resource-constrained cloud instances while retaining full Qwen3-VL capability set

Reach for something else if

• Highest-accuracy multimodal reasoning tasks — for MMMU, MathVision, and complex multi-step visual reasoning benchmarks, qwen/qwen3-vl-32b (76.0 MMMU Instruct) or qwen/qwen3-vl-235b provide meaningfully better results
• Pure text tasks with no visual input — qwen/qwen3-8b is the right choice for text-only workloads; the vision backbone adds overhead with no benefit
• Audio or native video generation — Qwen3-VL-8B is image+text-in / text-out only; it cannot produce audio, video, or rendered image outputs

Context Length

33K

Max Output

Input Priceper 1M tokens

$0.1944/ 1M tokens

Output Priceper 1M tokens

$0.756/ 1M tokens

Modalities

textimage→text

Capabilities

Vision

Pricing Breakdown

Type	Rate
Input	$0.1944 / 1M tokens
Output	$0.756 / 1M tokens

Supported Parameters

temperaturemax_tokenstop_pstop

Specifications

Release date	2025-10-15github.com ↗	verified
Architecture	Dense transformer; ~8B parameters; enhanced ViT image encoder with Interleaved-MRoPE and DeepStackarxiv.org ↗	verified
Input modalities	Text, images, video (interleaved sequences supported)huggingface.co ↗	verified
Native context length	256K tokens (expandable to 1M)github.com ↗	verified
OCR languages	32 languageshuggingface.co ↗	verified
License	Apache 2.0huggingface.co ↗	verified
Training cutoff	Not publicly disclosed	unknown

Benchmarks

Benchmark	Score	Source
DocVQA (test)	96.1%%	llm-stats.com ↗
ScreenSpot	94.4%%	llm-stats.com ↗
OCRBench	89.6%%	llm-stats.com ↗
AI2D	85.7%%	llm-stats.com ↗
MMBench-V1.1	85.0%%	llm-stats.com ↗

API Usage Examples

Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.

cURL

curl https://api.therouter.ai/v1/chat/completions   -H "Content-Type: application/json"   -H "Authorization: Bearer $THE_ROUTER_API_KEY"   -d '{
    "model": "qwen/qwen3-vl-8b",
    "messages": [
      {"role": "user", "content": "Summarize the key points from this input."}
    ]
  }'

API guide

Vision chat completion

cURL

curl https://api.therouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen/qwen3-vl-8b",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "image_url", "image_url": {"url": "https://therouter.ai/assets/vision-sample.png"}},
        {"type": "text", "text": "Describe what you see in this image."}
      ]
    }]
  }'

Similar models

Cross-provider sibling models

qwen/qwen3-vl-32b

Larger dense sibling in the Qwen3-VL family; choose when peak multimodal reasoning quality matters more than cost

qwen/qwen3-vl-235b

MoE flagship (22B active) for maximum quality at fleet scale; significantly higher cost but best-in-class multimodal benchmark scores

google/gemini-2.5-flash

Closed-model alternative with strong vision at competitive price; benchmark against Qwen3-VL-8B for your specific workload before choosing

News & changes

2025-11-27

Qwen3-VL technical report published on arXiv (2511.21631)

The Qwen Team released the full technical report covering the Interleaved-MRoPE positional encoding scheme, DeepStack vision-language alignment, text-timestamp video grounding, and benchmark results across the full Qwen3-VL family including the 8B dense model.

re-authored by TheRouterarxiv.org ↗

2025-10-15

Qwen3-VL-8B Instruct and Thinking editions released

Alibaba Qwen released the Qwen3-VL-8B in both Instruct and Thinking editions on Hugging Face, alongside the Qwen3-VL-4B. The 8B Instruct achieves DocVQA 96.1% and ScreenSpot 94.4%, matching much larger models; the Thinking edition adds chain-of-thought reasoning for STEM-heavy use cases.

re-authored by TheRoutergithub.com/QwenLM ↗

Frequently asked

When should I choose Qwen3-VL-8B over Qwen3-VL-32B?

Choose Qwen3-VL-8B when per-token cost or GPU memory footprint is a hard constraint, your workload is document OCR, GUI automation, or image QA (where the 8B scores are competitive with larger models), or you need to run multiple concurrent requests on a fixed-size instance. Choose Qwen3-VL-32B when benchmark accuracy on MMMU, MathVision, or complex multi-step visual reasoning matters more than cost.

github.com ↗

What is the difference between Qwen3-VL-8B Instruct and Thinking on TheRouter?

Instruct is optimised for direct responses — lower latency and token cost. Thinking produces step-by-step chain-of-thought reasoning before the final answer, increasing token count and cost but improving accuracy on math-embedded images and multi-step visual tasks. For document OCR, GUI tasks, and image QA, Instruct is usually the better choice. For STEM, competition math, or scientific diagram analysis, the Thinking edition closes the gap with larger models.

huggingface.co ↗

Can Qwen3-VL-8B process video, or only images?

Qwen3-VL-8B natively supports video input in addition to images and text, with a 256K-token context window expandable to 1M. It provides second-level temporal indexing and full recall across long videos via the Text–Timestamp Alignment architecture. Video support availability through TheRouter's /v1/chat/completions endpoint may differ from direct API access — check the current TheRouter model card for supported input types.