AI Index

Qwen3-VL-8B-Thinking

Language Alibaba Qwen Verified Released Sep 30, 2025
  • In TextImageVideo Out Text
  • Reasoning
  • Structured output
  • Open weights

Prices

3 offerings · prices as listed
Alibaba Cloud Model Studio Official International · USD qwen3-vl-8b-thinking
Conditions/ 1M tokensInputOutput
Standard no longer listed since Oct 11, 2026 $0.18$2.10

Prices from Alibaba Cloud Model Studio pricing (checked Oct 11, 2026)

Alibaba Cloud Model Studio Official Global · USD qwen3-vl-8b-thinking
Conditions/ 1M tokensInputOutput
Standard no longer listed since Oct 11, 2026 $0.072$0.717

Prices from Alibaba Cloud Model Studio pricing (checked Oct 11, 2026)

Alibaba Cloud Model Studio Official China · CNY qwen3-vl-8b-thinking
Conditions/ 1M tokensInputOutput
Standard no longer listed since Oct 11, 2026 ¥0.50¥5.00

Prices from 阿里云百炼：模型价格 (checked Oct 6, 2026)

Standard: the price a request with no conditions pays (the headline price elsewhere on the site).

Details

Max input
126.98K
Tool use
—
Family
qwen-vl-thinking
Model ID

Not stated on Alibaba Qwen’s pages: reasoning efforts, reasoning can be off, max thinking budget, built-in tools, knowledge cutoff, aliases.

Run locally

Weights: Qwen/Qwen3-VL-8B-Thinking
  • Parameters 8.8B
  • Dense
  • Native precision BF16
  • Layers 36
  • KV cache per token 144 KB
  • Config positions 256K (config.json max_position_embeddings)
  • API context 128K (maker’s API)
My hardware GeForce RTX 4090 24 GB · 1008 GB/s · 64 GB DDR5 dual channel · 89.6 GB/s
On your hardware · GeForce RTX 4090
Runs well BF16 (native) (automatic) 8K context 27–40 tok/sOne GPU
WeightsKV cacheOverheadUsable
Needs
≈ 18.3 GB (weights 16.3 (maker’s file) + KV cache 1.1 + overhead 0.8)
Usable
21.6 GB (90% of its memory)
Generation speed
27–40 tok/s
Longest context
31K (longer doesn’t fit)
First token
about 0.28–0.42 s for a 2K prompt
PrecisionWeightsNeeds at 8KOn your hardwareLongest context
BF16 (native)maker’s file16.3 GB≈ 18.3Runs well 27–40 tok/s31K
Q8_0maker’s file8.1 GB≈ 9.7Runs well 51–76 tok/s92K
FP8maker’s file9.9 GB≈ 11.5Runs well 43–64 tok/s80K
Q6_K6.7 GB≈ 8.3Runs well 60–90 tok/s102K
Q5_K_M5.8 GB≈ 7.4Runs well 68–102 tok/s109K
Q5_05.6 GB≈ 7.2Runs well 70–104 tok/s110K
Q4_K_Mmaker’s file4.7 GB≈ 6.3Runs well 81–121 tok/s117K
Q4_04.6 GB≈ 6.2Runs well 82–123 tok/s117K
IQ4_XS4.3 GB≈ 6.0Runs well 86–129 tok/s119K
Q3_K_Mquality drops4.0 GB≈ 5.6Runs well 92–138 tok/s122K
IQ3_XXSquality drops3.1 GB≈ 4.7Runs well 110–166 tok/s128K
Q2_Kquality drops3.1 GB≈ 4.7Runs well 112–168 tok/s128K
IQ2_XXSquality drops2.1 GB≈ 3.7Runs well 145–218 tok/s128K

Click a row to choose a precision. “Maker’s file” sizes are the maker’s own files; others are parameters × bits per weight.

Compared with other hardware

Each at the highest precision it runs. Click one for details.

Smallest that runs it (Q4_K_M, 8K): GPU GeForce RTX 3080; unified memory Apple M5 · 16 GB.

Run commands

vLLM
vllm serve Qwen/Qwen3-VL-8B-Thinking --max-model-len 8192
Memory to fine-tune
Full≈ 133.1 GBDoesn’t fit your hardware
LoRA≈ 19.0 GBFits your hardware
QLoRA≈ 7.3 GBFits your hardware

Batch 1, 2,048-token sequences, gradient checkpointing; LoRA and QLoRA train rank-16 adapters on the attention projections.

Estimates; facts from the maker’s Hugging Face repositories, as of 2026-11-06. How it’s estimated

Cited scores

From third-party leaderboards, not measured by us

This model isn’t on any leaderboard we cite yet.

Changes

All changes →

No changes recorded yet.

Sources

Alibaba Qwen’s official policies: data use, privacy, terms →

Alibaba Qwen · China · qwen.ai

Snapshots, the same family and other models from this provider