Instructions to use Zyphra/ZAYA1-VL-8B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Zyphra/ZAYA1-VL-8B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="Zyphra/ZAYA1-VL-8B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoModelForSeq2SeqLM
model = AutoModelForSeq2SeqLM.from_pretrained("Zyphra/ZAYA1-VL-8B", dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use Zyphra/ZAYA1-VL-8B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Zyphra/ZAYA1-VL-8B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Zyphra/ZAYA1-VL-8B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/Zyphra/ZAYA1-VL-8B

SGLang

How to use Zyphra/ZAYA1-VL-8B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Zyphra/ZAYA1-VL-8B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Zyphra/ZAYA1-VL-8B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Zyphra/ZAYA1-VL-8B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Zyphra/ZAYA1-VL-8B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use Zyphra/ZAYA1-VL-8B with Docker Model Runner:
```
docker model run hf.co/Zyphra/ZAYA1-VL-8B
```

hassansh commited on 15 days ago

Commit

d99ffdb

verified ·

1 Parent(s): 90eaeb1

Update README.md

Browse files

Files changed (1) hide show

README.md +27 -27

README.md CHANGED Viewed

@@ -7,16 +7,16 @@ pipeline_tag: image-text-to-text
 # ZAYA1-VL-8B
-ZAYA1-8B-VL is a vision-language model (VLM) built upon Zyphra's ZAYA1-7B LLM. It has state-of-the-art performance among VLMs for its size and inference efficiency.
 Learn more about our vision-language models in our [announcement blog post](-/TODO) and our [accompanying technical report](-/TODO)
-ZAYA1-8B-VL is open-sourced under the Apache 2.0 license.
 ## Performance
-ZAYA1-8B-VL performs extremely strongly against models of a comparable size and inference flops including outperforming several strong larger models.
 ![scores_vs_params-1](https://cdn-uploads.huggingface.co/production/uploads/65c05e75c084467acab2f84a/ivOwANwAmAzFDsa-JveP6.png)
@@ -33,38 +33,38 @@ ZAYA1-8B-VL performs extremely strongly against models of a comparable size and
 ### Model Architecture
-ZAYA1-8B-VL builds upon and uses our [ZAYA1-7B LLM](https://huggingface.co/Zyphra/ZAYA1-base) as its base text encoder. We also use the [Qwen2.5-VL vision encoder](https://huggingface.co/docs/transformers/model_doc/qwen2_5_vl) for the ViT. ZAYA1-8B-VL introduces two novel architectural innovations:
-* **Vision-specific LoRA parameters**: ZAYA1-8B-VL utilizes specialized LoRA parameters on its MLPs and CCA weights which are only activated on vision tokens. We find that adding vision-specific parameters substantially improves model performance since the model has the option to devote specific parameters solely to visual processing. We train these LoRA parameters alongside the main model parameters during training.
-* **Bidirectional Attention for image tokens**: ZAYA1-8B-VL processes all image token inputs with a bidirectional attention mask, meaning attention is not causal across an image. We find that this improves performance by not imposing an arbitrary causal order to image tokens which are naturally non-causal.
 ![smoe_vl_arch-1](https://cdn-uploads.huggingface.co/production/uploads/65c05e75c084467acab2f84a/wFuVyNIAzcaehBG1RiMjg.png)
-ZAYA1-8B-VL is trained only upon open data. Detailed dataset descriptions can be found in the accompanying technical report.
-| Eval | ZAYA1-VL-8B-A1B | MolmoE-8B-A1B | InternVL3.5-20B-A4B | Qwen3.5-2B | Molmo2-4B | Qwen3.5-4B |
 |---|---:|---:|---:|---:|---:|---:|
-| AI2D (test) | **87.5** | <u>73.6</u> | 85.5 | 78.6 | 85.4 | 83.7 |
-| ChartQA (test) | 82.2 | <u>77.9</u> | **87.0** | 78.4 | 86.1 | 82.4 |
-| DocVQA (test) | 92.5 | <u>77.7</u> | 92.9 | -- | 87.8 | -- |
-| InfoVQA (test) | 74.0 | <u>53.9</u> | 78.1 | -- | 78.6 | -- |
-| TextVQA (val) | <u>74.4</u> | 78.1 | 78.5 | 79.0 | **83.1** | 81.1 |
-| OCRBench | 79.8 | <u>55.0</u> | **86.7** | 83.1 | 62.0 | 85.3 |
-| VQA v2.0 (val) | 80.0 | 82.8 | 78.4 | 78.3 | **85.3** | 80.4 |
-| MathVista (mini) | 64.0 | <u>39.1</u> | 73.5 | 52.9 | 56.5 | **82.3** |
-| MMMU (val) | 46.0 | -- | **72.6** | 49.2 | <u>48.8</u> | 56.9 |
-| SEED (image) | 72.7 | <u>68.7</u> | 76.8 | 75.8 | **78.0** | 76.6 |
-| Blink (val) | <u>45.9</u> | -- | 58.9 | 61.0 | **63.5** | 56.8 |
-| RealWorldQA | 65.0 | <u>60.4</u> | 71.2 | 69.0 | 73.8 | **74.2** |
-| CountBenchQA | 88.1 | 77.4 | 82.1 | 84.2 | **91.2** | 84.8 |
-| PixMoCount (test) | 83.1 | <u>45.2</u> | 47.3 | 65.5 | **87.0** | 84.2 |
-| Point-Bench (avg) | 58.0 | 58.0 | -- | <u>40.6</u> | **68.5** | 64.4 |
-| RefCOCO (avg) | 84.3 | -- | **89.1** | <u>80.1</u> | -- | 87.7 |
-(based on VLMEvalKit)
 ## Quick start

 # ZAYA1-VL-8B
+ZAYA1-VL-8B is a vision-language model (VLM) built upon Zyphra's ZAYA1-7B LLM. It has state-of-the-art performance among VLMs for its size and inference efficiency.
 Learn more about our vision-language models in our [announcement blog post](-/TODO) and our [accompanying technical report](-/TODO)
+ZAYA1-VL-8B is open-sourced under the Apache 2.0 license.
 ## Performance
+ZAYA1-VL-8B performs extremely strongly against models of a comparable size and inference flops including outperforming several strong larger models.
 ![scores_vs_params-1](https://cdn-uploads.huggingface.co/production/uploads/65c05e75c084467acab2f84a/ivOwANwAmAzFDsa-JveP6.png)
 ### Model Architecture
+ZAYA1-VL-8B builds upon and uses our [ZAYA1-7B LLM](https://huggingface.co/Zyphra/ZAYA1-base) as its base text encoder. We also use the [Qwen2.5-VL vision encoder](https://huggingface.co/docs/transformers/model_doc/qwen2_5_vl) for the ViT. ZAYA1-VL-8B introduces two novel architectural innovations:
+* **Vision-specific LoRA parameters**: ZAYA1-VL-8B utilizes specialized LoRA parameters on its MLPs and CCA weights which are only activated on vision tokens. We find that adding vision-specific parameters substantially improves model performance since the model has the option to devote specific parameters solely to visual processing. We train these LoRA parameters alongside the main model parameters during training.
+* **Bidirectional Attention for image tokens**: ZAYA1-VL-8B processes all image token inputs with a bidirectional attention mask, meaning attention is not causal across an image. We find that this improves performance by not imposing an arbitrary causal order to image tokens which are naturally non-causal.
 ![smoe_vl_arch-1](https://cdn-uploads.huggingface.co/production/uploads/65c05e75c084467acab2f84a/wFuVyNIAzcaehBG1RiMjg.png)
+ZAYA1-VL-8B is trained only upon open data. Detailed dataset descriptions can be found in the accompanying technical report.
+| Eval | ZAYA1-VL-8B(0.7B / 8B) | MolmoE(1.2B / 8B) | Qwen3.5-2B | InternVL3.5-20B(20B / 4B) | Molmo2-4B | Qwen3.5-4B |
 |---|---:|---:|---:|---:|---:|---:|
+| AI2D (test) | **87.5** | <u>73.6</u> | 78.6 | 85.5 | 85.4 | 83.7 |
+| ChartQA (test) | 82.2 | <u>77.9</u> | 78.4 | **87.0** | 86.1 | 82.4 |
+| DocVQA (test) | 92.5 | <u>77.7</u> | -- | 92.9 | 87.8 | -- |
+| InfoVQA (test) | 74.0 | <u>53.9</u> | -- | 78.1 | 78.6 | -- |
+| TextVQA (val) | <u>74.4</u> | 78.1 | 79.0 | 78.5 | **83.1** | 81.1 |
+| OCRBench | 79.8 | <u>55.0</u> | 83.1 | **86.7** | 62.0 | 85.3 |
+| VQA v2.0 (val) | 80.0 | 82.8 | 78.3 | 78.4 | **85.3** | 80.4 |
+| MathVista (mini) | 64.0 | <u>39.1</u> | 52.9 | 73.5 | 56.5 | **82.3** |
+| MMMU (val) | 46.0 | -- | 49.2 | **72.6** | <u>48.8</u> | 56.9 |
+| SEED (image) | 72.7 | <u>68.7</u> | 75.8 | 76.8 | **78.0** | 76.6 |
+| Blink (val) | <u>45.9</u> | -- | 61.0 | 58.9 | **63.5** | 56.8 |
+| RealWorldQA | 65.0 | <u>60.4</u> | 69.0 | 71.2 | 73.8 | **74.2** |
+| CountBenchQA | 88.1 | 77.4 | 84.2 | 82.1 | **91.2** | 84.8 |
+| PixMoCount (test) | 83.1 | <u>45.2</u> | 65.5 | 47.3 | **87.0** | 84.2 |
+| Point-Bench (avg) | 58.0 | 58.0 | <u>40.6</u> | -- | **68.5** | 64.4 |
+| RefCOCO (avg) | 84.3 | -- | <u>80.1</u> | **89.1** | -- | 87.7 |
+All numbers are run on the Zyphra evaluation harness (based on VLMEvalKit). Models are ordered by total parameter count.
 ## Quick start