Web4/Linkspreed-MLM
15
1---2library_name: transformers.js3base_model: google/gemma-4-E2B-it4license: apache-2.05license_link: https://ai.google.dev/gemma/docs/gemma_4_license6tags:7- conversational8- image-text-to-text9pipeline_tag: any-to-any10---11<div align="center">12 <img src=https://ai.google.dev/gemma/images/gemma4_banner.png>13</div>14 15 16<p align="center">17 <a href="https://huggingface.co/collections/google/gemma-4" target="_blank">Hugging Face</a> |18 <a href="https://github.com/google-gemma" target="_blank">GitHub</a> |19 <a href="https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/" target="_blank">Launch Blog</a> |20 <a href="https://ai.google.dev/gemma/docs/core" target="_blank">Documentation</a>21 <br>22 <b>License</b>: <a href="https://ai.google.dev/gemma/docs/gemma_4_license" target="_blank">Apache 2.0</a> | <b>Authors</b>: <a href="https://deepmind.google/models/gemma/" target="_blank">Google DeepMind</a>23</p>24 25Gemma is a family of open models built by Google DeepMind. Gemma 4 models are multimodal, handling text and image input (with audio supported on small models) and generating text output. This release includes open-weights models in both pre-trained and instruction-tuned variants. Gemma 4 features a context window of up to 256K tokens and maintains multilingual support in over 140 languages. 26 27Featuring both Dense and Mixture-of-Experts (MoE) architectures, Gemma 4 is well-suited for tasks like text generation, coding, and reasoning. The models are available in four distinct sizes: **E2B**, **E4B**, **26B A4B**, and **31B**. Their diverse sizes make them deployable in environments ranging from high-end phones to laptops and servers, democratizing access to state-of-the-art AI.28 29Gemma 4 introduces key **capability and architectural advancements**:30 31* **Reasoning** – All models in the family are designed as highly capable reasoners, with configurable thinking modes.32 33* **Extended Multimodalities** – Processes Text, Image with variable aspect ratio and resolution support (all models), Video, and Audio (featured natively on the E2B and E4B models).34 35* **Diverse & Efficient Architectures** – Offers Dense and Mixture-of-Experts (MoE) variants of different sizes for scalable deployment.36 37* **Optimized for On-Device** – Smaller models are specifically designed for efficient local execution on laptops and mobile devices.38 39* **Increased Context Window** – The small models feature a 128K context window, while the medium models support 256K.40 41* **Enhanced Coding & Agentic Capabilities** – Achieves notable improvements in coding benchmarks alongside native function-calling support, powering highly capable autonomous agents.42 43* **Native System Prompt Support** – Gemma 4 introduces native support for the `system` role, enabling more structured and controllable conversations.44 45## **Models Overview**46 47Gemma 4 models are designed to deliver frontier-level performance at each size, targeting deployment scenarios from mobile and edge devices (E2B, E4B) to consumer GPUs and workstations (26B A4B, 31B). They are well-suited for reasoning, agentic workflows, coding, and multimodal understanding.48 49The models employ a hybrid attention mechanism that interleaves local sliding window attention with full global attention, ensuring the final layer is always global. This hybrid design delivers the processing speed and low memory footprint of a lightweight model without sacrificing the deep awareness required for complex, long-context tasks. To optimize memory for long contexts, global layers feature unified Keys and Values, and apply Proportional RoPE (p-RoPE). 50 51### Dense Models52 53| Property | E2B | E4B | 31B Dense |54| :---- | :---- | :---- | :---- |55| **Total Parameters** | 2.3B effective (5.1B with embeddings) | 4.5B effective (8B with embeddings) | 30.7B |56| **Layers** | 35 | 42 | 60 |57| **Sliding Window** | 512 tokens | 512 tokens | 1024 tokens |58| **Context Length** | 128K tokens | 128K tokens | 256K tokens |59| **Vocabulary Size** | 262K | 262K | 262K |60| **Supported Modalities** | Text, Image, Audio | Text, Image, Audio | Text, Image |61| **Vision Encoder Parameters** | *~150M* | *~150M* | *~550M* |62| **Audio Encoder Parameters** | *~300M* | *~300M* | No Audio |63 64The "E" in E2B and E4B stands for "effective" parameters. The smaller models incorporate Per-Layer Embeddings (PLE) to maximize parameter efficiency in on-device deployments. Rather than adding more layers or parameters to the model, PLE gives each decoder layer its own small embedding for every token. These embedding tables are large but are only used for quick lookups, which is why the effective parameter count is much smaller than the total.65 66### Mixture-of-Experts (MoE) Model67 68| Property | 26B A4B MoE |69| :---- | :---- |70| **Total Parameters** | 25.2B |71| **Active Parameters** | 3.8B |72| **Layers** | 30 |73| **Sliding Window** | 1024 tokens |74| **Context Length** | 256K tokens |75| **Vocabulary Size** | 262K |76| **Expert Count** | 8 active / 128 total and 1 shared |77| **Supported Modalities** | Text, Image |78| **Vision Encoder Parameters** | *~550M* |79 80The "A" in 26B A4B stands for "active parameters" in contrast to the total number of parameters the model contains. By only activating a 4B subset of parameters during inference, the Mixture-of-Experts model runs much faster than its 26B total might suggest. This makes it an excellent choice for fast inference compared to the dense 31B model since it runs almost as fast as a 4B-parameter model.81 82## **Benchmark Results** 83 84These models were evaluated against a large collection of different datasets and metrics to cover different aspects of text generation. Evaluation results marked in the table are for instruction-tuned models.85 86| | Gemma 4 31B | Gemma 4 26B A4B | Gemma 4 E4B | Gemma 4 E2B | Gemma 3 27B (no think) |87| :---- | :---- | :---- | :---- | :---- | :---- |88| MMLU Pro | 85.2% | 82.6% | 69.4% | 60.0% | 67.6% |89| AIME 2026 no tools | 89.2% | 88.3% | 42.5% | 37.5% | 20.8% |90| LiveCodeBench v6 | 80.0% | 77.1% | 52.0% | 44.0% | 29.1% |91| Codeforces ELO | 2150 | 1718 | 940 | 633 | 110 |92| GPQA Diamond | 84.3% | 82.3% | 58.6% | 43.4% | 42.4% |93| Tau2 (average over 3) | 76.9% | 68.2% | 42.2% | 24.5% | 16.2% |94| HLE no tools | 19.5% | 8.7% | - | - | - |95| HLE with search | 26.5% | 17.2% | - | - | - |96| BigBench Extra Hard | 74.4% | 64.8% | 33.1% | 21.9% | 19.3% |97| MMMLU | 88.4% | 86.3% | 76.6% | 67.4% | 70.7% |98| **Vision** | | | | | |99| MMMU Pro | 76.9% | 73.8% | 52.6% | 44.2% | 49.7% |100| OmniDocBench 1.5 (average edit distance, lower is better) | 0.131 | 0.149 | 0.181 | 0.290 | 0.365 |101| MATH-Vision | 85.6% | 82.4% | 59.5% | 52.4% | 46.0% |102| MedXPertQA MM | 61.3% | 58.1% | 28.7% | 23.5% | - |103| **Audio** | | | | | |104| CoVoST | - | - | 35.54 | 33.47 | - |105| FLEURS (lower is better) | - | - | 0.08 | 0.09 | - |106| **Long Context** | | | | | |107| MRCR v2 8 needle 128k (average) | 66.4% | 44.1% | 25.4% | 19.1% | 13.5% |108 109## **Core Capabilities**110 111Gemma 4 models handle a broad range of tasks across text, vision, and audio. Key capabilities include:112 113* **Thinking** – Built-in reasoning mode that lets the model think step-by-step before answering.114* **Long Context** – Context windows of up to 128K tokens (E2B/E4B) and 256K tokens (26B A4B/31B).115* **Image Understanding** – Object detection, Document/PDF parsing, screen and UI understanding, chart comprehension, OCR (including multilingual), handwriting recognition, and pointing. Images can be processed at variable aspect ratios and resolutions.116* **Video Understanding** – Analyze video by processing sequences of frames.117* **Interleaved Multimodal Input** – Freely mix text and images in any order within a single prompt.118* **Function Calling** – Native support for structured tool use, enabling agentic workflows.119* **Coding** – Code generation, completion, and correction.120* **Multilingual** – Out-of-the-box support for 35+ languages, pre-trained on 140+ languages.121* **Audio** (E2B and E4B only) – Automatic speech recognition (ASR) and speech-to-translated-text translation across multiple languages.122 123## **Best Practices**124 125For the best performance, use these configurations and best practices:126 127### 1. Sampling Parameters128 129Use the following standardized sampling configuration across all use cases:130 131* `temperature=1.0` 132* `top_p=0.95` 133* `top_k=64`134 135### 2. Thinking Mode Configuration136 137Compared to Gemma 3, the models use standard `system`, `assistant`, and `user` roles. To properly manage the thinking process, use the following control tokens:138 139* **Trigger Thinking:** Thinking is enabled by including the `<|think|>` token at the start of the system prompt. To disable thinking, remove the token. 140* **Standard Generation:** When thinking is enabled, the model will output its internal reasoning followed by the final answer using this structure: 141 `<|channel>thought\n`**[Internal reasoning]**`<channel|>` 142* **Disabled Thinking Behavior:** For all models except for the E2B and E4B variants, if thinking is disabled, the model will still generate the tags but with an empty thought block: 143 `<|channel>thought\n<channel|>`**[Final answer]**144 145> [!Note]146> Note that many libraries like Transformers and llama.cpp handle the complexities of the chat template for you.147 148### 3. Multi-Turn Conversations149 150* **No Thinking Content in History**: In multi-turn conversations, the historical model output should only include the final response. Thoughts from previous model turns must *not be added* before the next user turn begins.151 152### 4. Modality order153 154* For optimal performance with multimodal inputs, place image and/or audio content **before** the text in your prompt. 155 156### 5. Variable Image Resolution157 158Aside from variable aspect ratios, Gemma 4 supports variable image resolution through a configurable visual token budget, which controls how many tokens are used to represent an image. A higher token budget preserves more visual detail at the cost of additional compute, while a lower budget enables faster inference for tasks that don't require fine-grained understanding.159 160* The supported token budgets are: **70**, **140**, **280**, **560**, and **1120**. 161 * Use *lower budgets* for classification, captioning, or video understanding, where faster inference and processing many frames outweigh fine-grained detail. 162 * Use *higher budgets* for tasks like OCR, document parsing, or reading small text.163 164### 6. Audio165 166Use the following prompt structures for audio processing:167 168* **Audio Speech Recognition (ASR)**169 170```text171Transcribe the following speech segment in {LANGUAGE} into {LANGUAGE} text.172 173Follow these specific instructions for formatting the answer:174* Only output the transcription, with no newlines.175* When transcribing numbers, write the digits, i.e. write 1.7 and not one point seven, and write 3 instead of three.176```177 178* **Automatic Speech Translation (AST)**179 180```text181Transcribe the following speech segment in {SOURCE_LANGUAGE}, then translate it into {TARGET_LANGUAGE}.182When formatting the answer, first output the transcription in {SOURCE_LANGUAGE}, then one newline, then output the string '{TARGET_LANGUAGE}: ', then the translation in {TARGET_LANGUAGE}.183```184 185### 7. Audio and Video Length186 187All models support image inputs and can process videos as frames whereas the E2B and E4B models also support audio inputs. Audio supports a maximum length of 30 seconds. Video supports a maximum of 60 seconds assuming the images are processed at one frame per second.188 189## Usage190 191### Transformers.js (JavaScript)192 193If you haven't already, you can install the [Transformers.js](https://huggingface.co/docs/transformers.js) JavaScript library from [NPM](https://www.npmjs.com/package/@huggingface/transformers) using:194 195```sh196npm i @huggingface/transformers197```198 199You can then use the model as follows:200 201```js202import {203 AutoProcessor,204 Gemma4ForConditionalGeneration,205 TextStreamer,206 load_image,207 read_audio,208} from "@huggingface/transformers";209 210// Load processor and model211const model_id = "onnx-community/gemma-4-E2B-it-ONNX";212const processor = await AutoProcessor.from_pretrained(model_id);213const model = await Gemma4ForConditionalGeneration.from_pretrained(model_id, {214 dtype: "q4f16",215 device: "webgpu",216 progress_callback: (info) => {217 if (info.status === "progress_total") {218 // console.log(`Loading model: ${info.progress}%`);219 }220 },221});222 223// Prepare prompt224const messages = [225 {226 role: "user",227 content: [228 { type: "image" },229 { type: "audio" },230 {231 type: "text",232 text: "Describe this image in detail and transcribe this audio verbatim.",233 },234 ],235 },236];237const prompt = processor.apply_chat_template(messages, {238 enable_thinking: false,239 add_generation_prompt: true,240});241 242// Prepare inputs243const image = await load_image("https://huggingface.co/datasets/Xenova/transformers.js-docs/resolve/main/artemis.jpeg");244const audio = await read_audio("https://huggingface.co/datasets/Xenova/transformers.js-docs/resolve/main/jfk.wav");245const inputs = await processor(prompt, image, audio, {246 add_special_tokens: false,247});248 249// Generate output250const outputs = await model.generate({251 ...inputs,252 max_new_tokens: 512,253 do_sample: false,254 streamer: new TextStreamer(processor.tokenizer, {255 skip_prompt: true,256 skip_special_tokens: false,257 // callback_function: (text) => { /* Do something with the streamed output */ },258 }),259});260 261// Decode output262const decoded = processor.batch_decode(263 outputs.slice(null, [inputs.input_ids.dims.at(-1), null]),264 { skip_special_tokens: true },265);266console.log(decoded[0]);267```268 269### ONNX Runtime (Python)270 271```py272import onnxruntime273import numpy as np274from transformers import AutoConfig, AutoProcessor, GenerationConfig275import os276from huggingface_hub import snapshot_download277 278# 1. Load models279## Load config and processor280model_id = "onnx-community/gemma-4-E2B-it-ONNX"281processor = AutoProcessor.from_pretrained(model_id)282config = AutoConfig.from_pretrained(model_id)283generation_config = GenerationConfig.from_pretrained(model_id)284 285## Load sessions286audio_model = "onnx/audio_encoder_q4.onnx"287vision_model = "onnx/vision_encoder_q4.onnx"288embed_model = "onnx/embed_tokens_q4.onnx"289decoder_model = "onnx/decoder_model_merged_q4.onnx"290model_dir = snapshot_download(model_id, allow_patterns=[f"{audio_model}*", f"{vision_model}*", f"{embed_model}*", f"{decoder_model}*"])291audio_model_path = os.path.join(model_dir, audio_model)292vision_model_path = os.path.join(model_dir, vision_model)293embed_model_path = os.path.join(model_dir, embed_model)294decoder_model_path = os.path.join(model_dir, decoder_model)295 296providers = ['CPUExecutionProvider']297vision_session = onnxruntime.InferenceSession(vision_model_path, providers=providers)298audio_session = onnxruntime.InferenceSession(audio_model_path, providers=providers)299embed_session = onnxruntime.InferenceSession(embed_model_path, providers=providers)300decoder_session = onnxruntime.InferenceSession(decoder_model_path, providers=providers)301 302## Set config values303eos_token_id = generation_config.eos_token_id304image_token_id = config.image_token_id305audio_token_id = config.audio_token_id306 307# 2. Prepare inputs308## Create input messages309messages = [310 {311 "role": "user",312 "content": [313 {"type": "text", "text": "In detail, describe the following audio and image."},314 {"type": "audio", "audio": "https://huggingface.co/datasets/Xenova/transformers.js-docs/resolve/main/jfk.wav"},315 {"type": "image", "image": "https://huggingface.co/datasets/Xenova/transformers.js-docs/resolve/main/artemis.jpeg"},316 ],317 },318]319inputs = processor.apply_chat_template(320 messages,321 add_generation_prompt=True,322 tokenize=True,323 return_dict=True,324 return_tensors="pt",325)326input_ids = inputs["input_ids"].numpy()327attention_mask = inputs["attention_mask"].numpy()328position_ids = np.cumsum(attention_mask, axis=-1) - 1329 330pixel_values = inputs["pixel_values"].numpy() if "pixel_values" in inputs else None331pixel_position_ids = inputs["image_position_ids"].numpy() if "image_position_ids" in inputs else None332input_features = inputs["input_features"].numpy().astype(np.float32) if "input_features" in inputs else None333input_features_mask = inputs["input_features_mask"].numpy() if "input_features_mask" in inputs else None334 335## Prepare decoder inputs336batch_size = input_ids.shape[0]337num_logits_to_keep = np.array(1, dtype=np.int64)338past_key_values = {339 inp.name: np.zeros(340 [batch_size, inp.shape[1], 0, inp.shape[3]],341 dtype=np.float32 if inp.type == "tensor(float)" else np.float16,342 )343 for inp in decoder_session.get_inputs()344 if inp.name.startswith("past_key_values")345}346 347# 3. Generation loop348max_new_tokens = 1024349generated_tokens = np.array([[]], dtype=np.int64)350image_features = None351audio_features = None352for i in range(max_new_tokens):353 inputs_embeds, per_layer_inputs = embed_session.run(None, {"input_ids": input_ids})354 if image_features is None and pixel_values is not None:355 image_features = vision_session.run(["image_features"], {"pixel_values": pixel_values, "pixel_position_ids": pixel_position_ids})[0]356 mask = (input_ids == image_token_id).reshape(-1)357 flat_embeds = inputs_embeds.reshape(-1, inputs_embeds.shape[-1])358 flat_embeds[mask] = image_features359 inputs_embeds = flat_embeds.reshape(inputs_embeds.shape)360 361 if audio_features is None and input_features is not None and input_features_mask is not None:362 audio_features = audio_session.run(363 ["audio_features"],364 {"input_features": input_features, "input_features_mask": input_features_mask},365 )[0]366 mask = (input_ids == audio_token_id).reshape(-1)367 flat_embeds = inputs_embeds.reshape(-1, inputs_embeds.shape[-1])368 flat_embeds[mask] = audio_features369 inputs_embeds = flat_embeds.reshape(inputs_embeds.shape)370 371 logits, *present_key_values = decoder_session.run(None, dict(372 inputs_embeds=inputs_embeds,373 attention_mask=attention_mask,374 per_layer_inputs=per_layer_inputs,375 position_ids=position_ids,376 num_logits_to_keep=num_logits_to_keep,377 **past_key_values,378 ))379 380 ## Update values for next generation loop381 input_ids = logits[:, -1].argmax(-1, keepdims=True)382 attention_mask = np.concatenate([attention_mask, np.ones_like(input_ids)], axis=-1)383 position_ids = position_ids[:, -1:] + 1384 for j, key in enumerate(past_key_values):385 past_key_values[key] = present_key_values[j]386 387 generated_tokens = np.concatenate([generated_tokens, input_ids], axis=-1)388 if np.isin(input_ids, eos_token_id).any():389 break390 391 ## (Optional) Streaming392 print(processor.decode(input_ids[0]), end="", flush=True)393print()394 395# 4. Output result396print(processor.batch_decode(generated_tokens, skip_special_tokens=True)[0])397```398 399## **Model Data**400 401Data used for model training and how the data was processed.402 403### **Training Dataset**404 405Our pre-training dataset is a large-scale, diverse collection of data encompassing a wide range of domains and modalities, which includes web documents, code, images, audio, with a cutoff date of January 2025. Here are the key components:406 407* **Web Documents**: A diverse collection of web text ensures the model is exposed to a broad range of linguistic styles, topics, and vocabulary. The training dataset includes content in over 140 languages. 408* **Code**: Exposing the model to code helps it to learn the syntax and patterns of programming languages, which improves its ability to generate code and understand code-related questions. 409* **Mathematics**: Training on mathematical text helps the model learn logical reasoning, symbolic representation, and to address mathematical queries. 410* **Images**: A wide range of images enables the model to perform image analysis and visual data extraction tasks.411 412The combination of these diverse data sources is crucial for training a powerful multimodal model that can handle a wide variety of different tasks and data formats.413 414### **Data Preprocessing**415 416Here are the key data cleaning and filtering methods applied to the training data:417 418* **CSAM Filtering**: Rigorous CSAM (Child Sexual Abuse Material) filtering was applied at multiple stages in the data preparation process to ensure the exclusion of harmful and illegal content. 419* **Sensitive Data Filtering**: As part of making Gemma pre-trained models safe and reliable, automated techniques were used to filter out certain personal information and other sensitive data from training sets. 420* **Additional methods**: Filtering based on content quality and safety in line with [our policies](https://ai.google/static/documents/ai-responsibility-update-published-february-2025.pdf).421 422## **Ethics and Safety**423 424As open models become central to enterprise infrastructure, provenance and security are paramount. Developed by Google DeepMind, Gemma 4 undergoes the same rigorous safety evaluations as our proprietary Gemini models. 425 426### **Evaluation Approach**427 428Gemma 4 models were developed in partnership with internal safety and responsible AI teams. A range of automated as well as human evaluations were conducted to help improve model safety. These evaluations align with [Google’s AI principles](https://ai.google/principles/), as well as safety policies, which aim to prevent our generative AI models from generating harmful content, including:429 430* Content related to child sexual abuse material and exploitation 431* Dangerous content (e.g., promoting suicide, or instructing in activities that could cause real-world harm) 432* Sexually explicit content 433* Hate speech (e.g., dehumanizing members of protected groups) 434* Harassment (e.g., encouraging violence against people)435 436### **Evaluation Results**437 438For all areas of safety testing, we saw major improvements in all categories of content safety relative to previous Gemma models. Overall, Gemma 4 models significantly outperform Gemma 3 and 3n models in improving safety, while keeping unjustified refusals low. All testing was conducted without safety filters to evaluate the model capabilities and behaviors. For both text-to-text and image-to-text, and across all model sizes, the model produced minimal policy violations, and showed significant improvements over previous Gemma models' performance. 439 440## **Usage and Limitations**441 442These models have certain limitations that users should be aware of.443 444### **Intended Usage**445 446Multimodal models (capable of processing vision, language, and/or audio) have a wide range of applications across various industries and domains. The following list of potential uses is not comprehensive. The purpose of this list is to provide contextual information about the possible use-cases that the model creators considered as part of model training and development.447 448* **Content Creation and Communication** 449 * **Text Generation**: These models can be used to generate creative text formats such as poems, scripts, code, marketing copy, and email drafts. 450 * **Chatbots and Conversational AI**: Power conversational interfaces for customer service, virtual assistants, or interactive applications. 451 * **Text Summarization**: Generate concise summaries of a text corpus, research papers, or reports. 452 * **Image Data Extraction**: These models can be used to extract, interpret, and summarize visual data for text communications. 453 * **Audio Processing and Interaction**: The smaller models (E2B and E4B) can analyze and interpret audio inputs, enabling voice-driven interactions and transcriptions. 454* **Research and Education** 455 * **Natural Language Processing (NLP) and VLM Research**: These models can serve as a foundation for researchers to experiment with VLM and NLP techniques, develop algorithms, and contribute to the advancement of the field. 456 * **Language Learning Tools**: Support interactive language learning experiences, aiding in grammar correction or providing writing practice. 457 * **Knowledge Exploration**: Assist researchers in exploring large bodies of text by generating summaries or answering questions about specific topics.458 459### **Limitations**460 461* **Training Data** 462 * The quality and diversity of the training data significantly influence the model's capabilities. Biases or gaps in the training data can lead to limitations in the model's responses. 463 * The scope of the training dataset determines the subject areas the model can handle effectively. 464* **Context and Task Complexity** 465 * Models perform well on tasks that can be framed with clear prompts and instructions. Open-ended or highly complex tasks might be challenging. 466 * A model's performance can be influenced by the amount of context provided (longer context generally leads to better outputs, up to a certain point). 467* **Language Ambiguity and Nuance** 468 * Natural language is inherently complex. Models might struggle to grasp subtle nuances, sarcasm, or figurative language. 469* **Factual Accuracy** 470 * Models generate responses based on information they learned from their training datasets, but they are not knowledge bases. They may generate incorrect or outdated factual statements. 471* **Common Sense** 472 * Models rely on statistical patterns in language. They might lack the ability to apply common sense reasoning in certain situations.473 474### **Ethical Considerations and Risks**475 476The development of vision-language models (VLMs) raises several ethical concerns. In creating an open model, we have carefully considered the following:477 478* **Bias and Fairness** 479 * VLMs trained on large-scale, real-world text and image data can reflect socio-cultural biases embedded in the training material. Gemma 4 models underwent careful scrutiny, input data pre-processing, and post-training evaluations as reported in this card to help mitigate the risk of these biases. 480* **Misinformation and Misuse** 481 * VLMs can be misused to generate text that is false, misleading, or harmful. 482 * Guidelines are provided for responsible use with the model, see the [Responsible Generative AI Toolkit](https://ai.google.dev/responsible). 483* **Transparency and Accountability** 484 * This model card summarizes details on the models' architecture, capabilities, limitations, and evaluation processes. 485 * A responsibly developed open model offers the opportunity to share innovation by making VLM technology accessible to developers and researchers across the AI ecosystem.486 487**Risks identified and mitigations**:488 489* **Generation of harmful content**: Mechanisms and guidelines for content safety are essential. Developers are encouraged to exercise caution and implement appropriate content safety safeguards based on their specific product policies and application use cases. 490* **Misuse for malicious purposes**: Technical limitations and developer and end-user education can help mitigate against malicious applications of VLMs. Educational resources and reporting mechanisms for users to flag misuse are provided. 491* **Privacy violations**: Models were trained on data filtered for removal of certain personal information and other sensitive data. Developers are encouraged to adhere to privacy regulations with privacy-preserving techniques. 492* **Perpetuation of biases**: It's encouraged to perform continuous monitoring (using evaluation metrics, human review) and the exploration of de-biasing techniques during model training, fine-tuning, and other use cases.493 494### **Benefits**495 496At the time of release, this family of models provides high-performance open vision-language model implementations designed from the ground up for responsible AI development compared to similarly sized models.