BLIP-2 accumulated 6,617 citations on Semantic Scholar as of May 2026 — making it one of the most-cited AI papers published in 2023. Salesforce Research’s vision-language model continues to attract over 500,000 monthly downloads on Hugging Face more than three years after release. This article covers the latest BLIP-2 statistics for 2026, including download trends, citation data, benchmark performance, model variants, and memory requirements.
BLIP-2 Statistics 2026 — TL;DR
BLIP-2’s Q-Former trains just 188 million parameters — under 2% of total model size — yet outperforms models 54 times larger on visual question answering tasks.
The blip2-opt-2.7b variant recorded 536,142 monthly downloads on Hugging Face as of December 2025.
Semantic Scholar lists 6,617 total citations for the BLIP-2 paper, with 863 classified as highly influential.
BLIP-2 scored 65.0% on zero-shot VQAv2, beating Flamingo80B’s 56.3% by 8.7 percentage points.
The model can run on consumer hardware with int4 quantization, requiring only 1.8 GB of memory.
Salesforce’s LAVIS library, which hosts BLIP-2, has collected over 10,400 GitHub stars.
How Many Downloads Does BLIP-2 Get on Hugging Face?
The blip2-opt-2.7b model — the most popular variant — recorded 536,142 monthly downloads on Hugging Face as of December 2025. That figure has remained steady since mid-2024, indicating consistent developer use rather than a one-time surge. The model page has 425 community likes, with over 100 Hugging Face Spaces running BLIP-2 for production and demo applications.
Developers have built 38 adapter models on top of BLIP-2’s architecture, along with 13 fine-tuned variants for domain-specific tasks. The broader BLIP model family on Hugging Face includes variants with download counts ranging from a few thousand to nearly 500,000 per month.
| Model Variant | Total Parameters | Monthly Downloads |
|---|---|---|
| blip2-opt-2.7b | ~4B | 491,000+ |
| blip2-opt-2.7b-coco | ~4B | 86,900+ |
| blip2-opt-6.7b | ~8B | 29,500+ |
| blip2-flan-t5-xl | ~8B | 6,700+ |
| blip2-flan-t5-xxl | ~12B | 6,700+ |
Source: Hugging Face Model Hub (Salesforce BLIP2 Collection, February 2025)
BLIP-2 Citation Statistics in 2026
Semantic Scholar recorded 6,617 total citations for BLIP-2 as of May 2026, up from 6,423 reported in December 2025. Of those, 863 are classified as highly influential — meaning they directly build on or extend BLIP-2’s methods. Another 1,266 appear in methods sections, while 1,050 show up in background sections of citing papers. The paper passed 3,000 citations within its first year of publication, placing it among the most-referenced AI research papers of 2023.
The combined BLIP series (BLIP and BLIP-2) has surpassed 15,000 total citations. New papers citing BLIP-2 continued appearing in 2026, including work presented at the ACM Web Conference in April 2026.
| Citation Category | Count |
|---|---|
| Total Citations | 6,617 |
| Highly Influential | 863 |
| Methods Citations | 1,266 |
| Background Citations | 1,050 |
| Results Citations | 31 |
Source: Semantic Scholar (May 2026)
How Does BLIP-2 Architecture Work?
BLIP-2 connects a frozen image encoder (EVA-CLIP ViT-g/14) to a frozen large language model through a lightweight Querying Transformer, or Q-Former. The Q-Former has 12 transformer layers that process 32 query tokens, each with 768-dimensional embeddings. This component contains 188 million trainable parameters — less than 2% of the total model when paired with an 11-billion-parameter language model.
Training happens in two stages. Stage one aligns visual and language representations using the frozen image encoder. Stage two connects the Q-Former output to a frozen LLM for text generation. The approach reduces computational requirements by roughly 98% compared to training the full model end-to-end. Three training objectives run simultaneously: image-text contrastive learning, image-text matching, and image-grounded text generation.
BLIP-2 Benchmark Performance
BLIP-2 scored 65.0% on zero-shot VQAv2, compared to Flamingo80B’s 56.3% — an 8.7 percentage point gap. On NoCaps captioning, it reached 121.6 CIDEr, beating the previous best of 113.2. The model also recorded 52.3% accuracy on GQA and 92.9% R@1 on Flickr30K for image-to-text retrieval.
These results came with 54 times fewer trainable parameters than Flamingo80B. Single GPU inference takes about one second per image on the OPT-2.7B variant, making BLIP-2 practical for real-time applications without specialized hardware clusters.
| Benchmark | BLIP-2 Score | Previous Best |
|---|---|---|
| Zero-shot VQAv2 | 65.0% | 56.3% (Flamingo80B) |
| NoCaps CIDEr | 121.6 | 113.2 |
| GQA Accuracy | 52.3% | — |
| Flickr30K R@1 (Image→Text) | 92.9% | — |
| COCO Caption CIDEr | 125.9 | — |
Source: BLIP-2 Paper (ICML 2023), Salesforce Research
BLIP-2 Memory Requirements By Precision
Hardware needs vary widely depending on numerical precision. The blip2-opt-2.7b variant takes 14.43 GB in float32 for inference — a number that drops to just 1.8 GB with int4 quantization through Bitsandbytes. Training with Adam optimizer multiplies these figures by four across all precision levels.
The int4 option is what makes BLIP-2 accessible on consumer GPUs. A standard 8 GB graphics card can comfortably handle int8 inference, while int4 fits on even lower-end hardware. This range of deployment options has helped maintain steady download numbers even as newer vision-language models have appeared.
| Precision | Model Size (Inference) | Training (Adam) |
|---|---|---|
| float32 | 14.43 GB | 57.72 GB |
| float16 / bfloat16 | 7.21 GB | 28.84 GB |
| int8 | 3.61 GB | 14.44 GB |
| int4 | 1.80 GB | 7.20 GB |
Source: Hugging Face Transformers Documentation
Where Does BLIP-2 Fit Among Vision-Language Models?
BLIP-2 arrived in January 2023 and was followed within months by LLaVA, MiniGPT-4, and InstructBLIP (Salesforce’s own instruction-tuned extension). Newer models like Qwen2.5-VL, LLaMA 3.2-Vision, and PaliGemma 2 have shifted toward simpler MLP adapters instead of BLIP-2’s Q-Former approach. That said, the Q-Former design still appears in active research — 1,266 methods citations show that researchers continue building on the architecture.
LLaVA outperformed BLIP-2 on conversational instruction-following tasks in direct comparisons, though BLIP-2 was not fine-tuned on instruction-following data. InstructBLIP addressed this gap, reaching state-of-the-art zero-shot results across 13 held-out datasets while keeping the Q-Former backbone intact. BLIP-2 remains the go-to reference model in vision-language research for parameter efficiency benchmarking.
BLIP-2 Developer Ecosystem in 2026
The LAVIS library — Salesforce’s open-source framework that includes BLIP-2 — has over 10,400 GitHub stars. BLIP-2 is also fully integrated into Hugging Face Transformers, where it can be loaded in three lines of Python code. The Hugging Face model hub lists 129 tagged BLIP-2 models, including community-built derivatives for image captioning, video understanding, medical imaging, and deepfake detection.
Beyond Hugging Face, BLIP-2 runs on platforms like Clarifai and Replicate. The M4-BLIP pipeline, published in December 2025, used BLIP-2 as a backbone for multi-modal manipulation detection, achieving 94.10 AUC in deepfake identification tasks. These downstream applications show the model’s reach well beyond standard chatbot and text generation use cases.
FAQ
How many downloads does BLIP-2 get per month?
The blip2-opt-2.7b variant recorded 536,142 monthly downloads on Hugging Face as of December 2025. Other variants add tens of thousands more.
How many citations does BLIP-2 have?
Semantic Scholar lists 6,617 total citations as of May 2026. Of those, 863 are highly influential and 1,266 appear in methods sections of citing papers.
What accuracy does BLIP-2 achieve on VQAv2?
BLIP-2 scored 65.0% on zero-shot VQAv2, outperforming Flamingo80B (56.3%) by 8.7 percentage points while using 54 times fewer trainable parameters.
How much memory does BLIP-2 need to run?
The OPT-2.7B variant requires 14.43 GB in float32, dropping to 1.8 GB with int4 quantization. Consumer GPUs with 8 GB can handle int8 inference.
Who created BLIP-2?
Salesforce Research released BLIP-2 in January 2023. The paper was authored by Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi, and presented at ICML 2023.
Sources:
https://huggingface.co/docs/transformers/en/model_doc/blip-2
