AxForge
Sign in Get started

Model reference · open weights

Qianfan-VL

LLMs baidu Vision + text 3 builds Its own licence terms 199 dl/mo

Qianfan-VL is an open-weight language model from Baidu. Qianfan-VL-8B (BF16) weighs 17.6 GB; the smallest configuration that runs it is 2× RTX 3060 12 GB.

  • Qianfan-VL is a series of general-purpose multimodal large language models developed by Baidu for enterprise-level applications, with a focus on document understanding and OCR.
  • The Qianfan-VL-3B variant contains 3.7B parameters, supports a 32k context length, and is designed for edge deployment and real-time OCR tasks.
  • It supports English and Chinese languages and is released under an other license.

Summary of the baidu/Qianfan-VL-3B model card, 2026-10-01. The estimate below is for another build of the family.

Hugging Face ↗GitHub ↗Paper ↗

What it is

Released byBaidu
Released2025-09-16
Parameters3.7B
VRAM17.6 GB for the weights

What it runs on

Memory and cards for Qianfan-VL-8B (BF16)

17.6 GBweights, file size
131 MBcache per 1K tokens
2.1 GBruntime overhead, at least
32,768 tokenscontext max
CardRequests at onceContext maxMemory
8K each32K each
RTX 3060 12 GB … RTX 4060 Ti 16 GB
2 smaller cards		———
RTX 3090 24 GB3—27K23.4 GB
RTX 4090 24 GB3—27K23.4 GB
RTX 5090 32 GB102all 32K31.0 GB
L40S 48 GB225all 32K44.0 GB
A100 80 GB5413all 32K78.2 GB
H100 80 GB4912all 32K78.1 GB
RTX PRO 6000 Blackwell 96 GB6416all 32K93.8 GB
DGX Spark (GB10) 128 GB unified7719all 32K107 GB
H200 141 GB10526all 32K138 GB
B200 180 GB14035all 32K176 GB
2× RTX 3060 12 GB
tensor parallel		1—10K11.6 GB a card
2× RTX 4060 Ti 16 GB
tensor parallel		82all 32K15.4 GB a card
2× RTX 4090 24 GB
tensor parallel		235all 32K23.4 GB a card
2× RTX 3090 24 GB
tensor parallel		235all 32K23.4 GB a card
Memory needed at each load
Requests at once8K tokens each32K tokens each
120.8 GB24.0 GB
525.1 GB41.2 GB
828.3 GB54.1 GB
1636.9 GB88.4 GB
3254.1 GB157 GB
6488.4 GB295 GB

One card, with vLLM's small-card settings.

Builds

Sizes, precisions & builds

BuildParamsPrecisionWeightsSmallest setup
Qianfan-VL-3B ↗ 3.7BBF16 7.4 GBRTX 3060 12 GB
Qianfan-VL-8B (above) ↗ 8.8BBF16 17.6 GB2× RTX 3060 12 GB
Qianfan-VL-70B ↗ 71.8BBF16 144 GB4× L40S 48 GB

From the model card

What Baidu says about Qianfan-VL

Read the model card

Domain Capability Enhancement through Continuous Pre-training | 3B to 70B Parameter Scale | Document Understanding & OCR Enhancement | Chain-of-Thought Reasoning Support

🔗 Quick Links

Model Description

Qianfan-VL is a series of general-purpose multimodal large language models enhanced for enterprise-level multimodal applications. The models offer deep optimization for high-frequency scenarios in industrial deployment while maintaining strong general capabilities.

Model Variants

ModelParametersContext LengthCoT SupportBest For
Qianfan-VL-3B3B32k❌Edge deployment, real-time OCR
Qianfan-VL-8B8B32k✅Server-side general scenarios, fine-tuning
Qianfan-VL-70B70B32k✅Complex reasoning, data synthesis

Architecture

  • Language Model:
    • Qianfan-VL-3B: Based on Qwen2.5-3B
    • Qianfan-VL-8B/70B: Based on Llama 3.1 architecture
    • Enhanced with 3T multilingual corpus
  • Vision Encoder: InternViT-based, supports dynamic patching up to 4K resolution
  • Cross-modal Fusion: MLP adapter for efficient vision-language bridging

Key Capabilities

🔍 OCR & Document Understanding

  • Full-Scenario OCR: Handwriting, formulas, natural scenes, cards/documents
  • Document Intelligence: Layout analysis, table parsing, chart understanding, document Q&A
  • High Precision: Industry-leading performance on OCR benchmarks

🧮 Chain-of-Thought Reasoning (8B & 70B)

  • Complex chart analysis and reasoning
  • Mathematical problem-solving with step-by-step derivation
  • Visual reasoning and logical inference
  • Statistical computation and trend prediction

📊 Benchmark Performance

General Vision-Language Benchmarks
BenchmarkQianfan-VL-3BQianfan-VL-8BQianfan-VL-70BInternVL-3-8BInternVL-3-78BQwen2.5-VL-7BQwen2.5-VL-72B
A-Bench_VAL75.6575.7278.175.8675.8676.4979.22
CCBench66.8670.3980.9877.8470.7857.6573.73
SEEDBench_IMG76.5578.0279.1377.077.5276.9878.34
SEEDBench2_Plus67.5970.9773.1769.5268.4770.9373.25
MMVet48.1753.2167.3480.2878.970.6475.69
MMMU_VAL46.4447.1158.3356.1160.7851.065.78
ScienceQA_TEST95.1997.6298.7697.9797.1785.4792.51
ScienceQA_VAL93.8597.6298.8197.8195.1483.5991.32
MMT-Bench_VAL62.2363.2271.0665.1763.6761.469.49
MTVQA_TEST26.530.1432.1830.327.6229.0831.48
BLINK49.9756.8159.4455.8751.8754.5563.02
MMStar57.9364.0769.4768.466.0761.5366.0
RealWorldQA65.7570.5971.6371.1174.2569.2873.86
Q-Bench1_VAL73.5175.2577.4675.9977.9978.179.93
POPE85.0886.0688.9790.5988.8785.9783.35
RefCOCO (Avg)85.9489.3791.0189.6591.4086.5690.25
OCR & Document Understanding
BenchmarkQianfan-VL-3BQianfan-VL-8BQianfan-VL-70BInternVL-3-8BInternVL-3-78BQwen2.5-VL-3BQwen2.5-VL-7BQwen2.5-VL-72B
OCRBench831854873881847810883874
AI2D_TEST81.3885.0787.2385.0783.5577.0780.47283.84
OCRVQA_TEST66.1568.9874.0639.0335.5869.2471.0266.8
TextVQA_VAL80.1182.1384.4882.1583.5279.0984.96283.26
DocVQA_VAL90.8593.5494.7592.0483.8292.71**9

Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.

© 2026 AxForge · EU-hosted AI infrastructure Pricing Docs Trust Privacy Terms