Skip to content
NEWLive updates feed from Hugging Face, arXiv and GitHub  See what changed →

Alibaba · Qwen2.5-VL

Qwen2.5-VL-32B-Instruct

An Apache-2.0 vision-language model that reads scanned documents, forms and IDs; the usual pick when OCR-free extraction is required.

Apache-2.0Dense128K contextVision

1,121,845 downloads · 501 likes · Model card · GitHub · synced 2026-09-19

Specifications

Parameters
33.5B
Architecture
dense · qwen2_5_vl
Layers
64
Hidden size
5,120
KV heads · head dim
8 · 128
Vocabulary
152,064
Native dtype
bfloat16
Context window
128,000 tokens
Released
2025-03-21
Last modified on Hub
2025-04-14

Features & licensing

License
Apache-2.0
Commercial use
Yes
Modalities
text, vision
Tool / function calling
No
Reasoning mode
No
Languages
EN, ZH, AR, FR, DE, ES
Quantised variants on Hub
AWQ, EXL2, FP8, GGUF, bitsandbytes
Pipeline
image-text-to-text
transformerssafetensorsqwen2_5_vlimage-text-to-textmultimodalconversationaleneval-resultstext-generation-inferenceendpoints_compatible

Task fit

Editorial scores (0–100) used by the recommendation engine.

  • Chat80
  • RAG / Q&A78
  • Code70
  • Summarisation78
  • Extraction88
  • Agents72

Hardware to run Qwen2.5-VL-32B-Instruct

Weights need about 74 GB at FP16, 39 GB at INT8 and 22 GB at INT4. The KV cache adds roughly 262 MB per 1,000 tokens per request. For a reference workload of 5 requests per second with 1,500 input and 300 output tokens, total GPU memory is around 36.2 GB, which fits on 1 × NVIDIA A100 (80 GB).

GPU configurations at the reference workload
ConfigurationVRAMUtilisationEst. first tokenCloud / month
2 × NVIDIA L4 (24 GB)48 GB75 %~0.5 s$1,168
1 × NVIDIA L40S (48 GB)48 GB75 %~0.2 s$1,387
1 × NVIDIA A100 (80 GB)recommended80 GB45 %~0.1 s$2,336
1 × NVIDIA H100 (80 GB)80 GB45 %~0.0 s$3,285

Estimates only. Run an assessment for your own traffic.

Use a vision-language model when inputs are images or scans. Budget extra memory for image tokens in the KV cache.