> For the complete documentation index, see [llms.txt](https://ultrasafe.gitbook.io/docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://ultrasafe.gitbook.io/docs/publish-your-docs.md).

# Models

### Overview

UltraSafe AI provides three types of models: advanced generalist models, specialized industry models, and cutting-edge research models.

### Advanced Generalist Models

<table data-header-hidden><thead><tr><th width="128">Model</th><th width="211">Weight availability</th><th width="59">Available via API</th><th width="337">Description</th><th width="74">Max Tokens</th><th>API Endpoints</th><th>Version</th></tr></thead><tbody><tr><td>EUS1</td><td>✔️ UltraSafe AI License</td><td>✔️</td><td>Our flagship model with state-of-the-art reasoning, knowledge, and coding capabilities. Ideal for complex tasks requiring extensive reasoning or specialized applications (Synthetic Text Generation, Code Generation, RAG, or Agents).</td><td>128k</td><td>eus1-latest</td><td>24.07</td></tr><tr><td>EUS Mega</td><td>✔️ UltraSafe AI License</td><td>✔️</td><td>Advanced multimodal AI model for complex tasks with extensive context understanding.</td><td>128k</td><td>eus-mega-latest</td><td>24.07</td></tr><tr><td>EUS Flash</td><td>✔️ UltraSafe AI License</td><td>✔️</td><td>High-speed text model for rapid processing, efficient for small-scale applications.</td><td>32k</td><td>eus-flash-latest</td><td>24.05</td></tr></tbody></table>

### Specialized Industry Models

<table data-header-hidden><thead><tr><th width="160">Model</th><th width="213">Weight availability</th><th width="54">Available via API</th><th width="316">Description</th><th width="65">Max Tokens</th><th>API Endpoints</th><th>Version</th></tr></thead><tbody><tr><td>UltraSafe AI - Healthcare</td><td>✔️ UltraSafe AI License</td><td>✔️</td><td>Specialized model for healthcare applications, optimized for medical terminology and reasoning.</td><td>64k</td><td>ultrasafe-healthcare</td><td>24.05</td></tr><tr><td>UltraSafe AI - Finance</td><td>✔️ UltraSafe AI License</td><td>✔️</td><td>Tailored for financial applications, with strong capabilities in numerical analysis and financial forecasting.</td><td>64k</td><td>ultrasafe-finance</td><td>24.05</td></tr><tr><td>UltraSafe AI - Code</td><td>✔️ UltraSafe AI License</td><td>✔️</td><td>A cutting-edge model optimized for code generation tasks, including fill-in-the-middle and code completion.</td><td>32k</td><td>ultrasafe-code</td><td>24.05</td></tr><tr><td>UltraSafe AI - Conversational</td><td>✔️ UltraSafe AI License</td><td>✔️</td><td>Specialized in natural language conversations, ideal for chatbots and virtual assistants.</td><td>32k</td><td>ultrasafe-conversational</td><td>24.05</td></tr><tr><td>UltraSafe AI - Education</td><td>✔️ UltraSafe AI License</td><td>✔️</td><td>Designed for educational applications, with strong capabilities in explanation and tutoring.</td><td>64k</td><td>ultrasafe-education</td><td>24.05</td></tr><tr><td>UltraSafe AI - Media</td><td>✔️ UltraSafe AI License</td><td>✔️</td><td>Optimized for media-related tasks, including content creation and analysis.</td><td>32k</td><td>ultrasafe-media</td><td>24.05</td></tr><tr><td>UltraSafe AI - Legal</td><td>✔️ UltraSafe AI License</td><td>✔️</td><td>Specialized in legal terminology and reasoning, ideal for legal document analysis and generation.</td><td>64k</td><td>ultrasafe-legal</td><td>24.05</td></tr><tr><td>UltraSafe AI - Cyber Security</td><td>✔️ UltraSafe AI License</td><td>✔️</td><td>Focused on cybersecurity applications, with capabilities in threat detection and security analysis.</td><td>32k</td><td>ultrasafe-cybersecurity</td><td>24.05</td></tr><tr><td>UltraSafe AI - Translation Engine</td><td>✔️ UltraSafe AI License</td><td>✔️</td><td>Specialized in multilingual translation, supporting over 100 languages.</td><td>32k</td><td>ultrasafe-translation</td><td>24.05</td></tr></tbody></table>

### Research Models

<table data-header-hidden><thead><tr><th width="159">Model</th><th width="196">Weight availability</th><th width="54">Available via API</th><th width="240">Description</th><th width="125">Max Tokens</th><th width="143">API Endpoints</th><th>Version</th></tr></thead><tbody><tr><td>UltraSafe Document Analyzer</td><td>✔️ Research License</td><td>✔️</td><td>Specialized in document analysis and information extraction.</td><td>64k</td><td>ultrasafe-doc-analyzer</td><td>v0.1</td></tr><tr><td>UltraSafe Data Reranker</td><td>✔️ Research License</td><td>✔️</td><td>Optimized for reranking and prioritizing data sets.</td><td>32k</td><td>ultrasafe-data-reranker</td><td>v0.1</td></tr><tr><td>UltraSafe Image Generation</td><td>✔️ Research License</td><td>✔️</td><td>Cutting-edge model for AI-powered image generation.</td><td>32k</td><td>ultrasafe-image-gen</td><td>v0.1</td></tr></tbody></table>

### Pricing

Please refer to the [pricing page](https://pricing.us.inc) for detailed information on costs.

### API Versioning

UltraSafe AI APIs are versioned with specific release dates. To prevent disruptions due to model updates and breaking changes, it is recommended to use the dated versions of the UltraSafe AI API. Be prepared for the deprecation of certain endpoints in the coming months.

Here are the details of the available versions:

* eus1-latest: currently points to eus1-2407.
* eus-mega-latest: currently points to eus-mega-2407.
* eus-flash-latest: currently points to eus-flash-2405.
* All specialized industry models are currently on version 24.05.

### Benchmark Results

UltraSafe AI models consistently rank among the top performers in industry-standard benchmarks. Our models offer top-tier reasoning capabilities and excel in multilingual tasks and code generation.

Key benchmark results:

* EUS1: Achieves 82.3% on MMLU (5-shot), 85.4% on HumanEval (code), and 89.7% on GSM8K (Math).
* EUS Mega: Outperforms EUS1 on complex reasoning tasks and multimodal applications.
* Specialized models: Each specialized model outperforms general models in its specific domain.

#### Choosing the Right Model

This guide will help you navigate the performance and cost considerations to select the most suitable UltraSafe AI model for your specific use cases.

UltraSafe AI models are the backbone of numerous large-scale applications. Below is a quick reference for matching use cases with the appropriate UltraSafe AI models:

* **Simple Tasks**: Tasks like classification, customer support, or basic text generation are best handled by the EUS Flash model.
* **Intermediate Tasks**: For tasks that require moderate reasoning, such as data extraction, summarization, or email writing, you can rely on specialized industry models or the EUS1 model.
* **Complex Tasks**: For more advanced tasks that require significant reasoning capabilities or are highly specialized, such as synthetic text generation, code generation, retrieval-augmented generation (RAG), or intelligent agents, the EUS1 or EUS Mega models are ideal choices.

### Performance and Cost Trade-offs

When selecting a model, evaluate performance and cost trade-offs based on your application's requirements. Generally, larger models offer better performance at a higher cost.

Performance ranking on the MMLU benchmark:

EUS1 (82.3%) > EUS Mega (80.1%) > Specialized Models (varies by domain) > EUS Flash (72.5%)

EUS1 currently outperforms other models across most benchmarks.

For independent benchmarks, refer to industry-standard evaluations to gain more insight into the performance and speed of different large language models.

As a general guideline, if you're new to working with Large Language Models, begin with EUS1 to assess its cost-effectiveness and performance. This will help you decide whether to transition to a specialized model or EUS Flashdepending on your specific requirements. If you're looking to replace an existing high-performance language model, EUS1or EUS Mega would likely be the optimal choice.

#### Use Cases

**EUS Flash: Efficient Handling of Simple, High-Volume Tasks**

EUS Flash is the ideal model for handling straightforward tasks that can be performed in bulk, such as classification, customer support, or basic text generation. It delivers strong performance at a cost-effective price point. For example, it can be effectively used for email classification to determine whether an email is spam:

#### Use Cases

**EUS Flash: Ideal for Simple, High-Volume Tasks**

EUS Flash is perfect for handling straightforward tasks that can be processed in large quantities, such as classification, customer support, or basic text generation. It delivers strong performance at a cost-effective price point. For example, EUS Flash can be efficiently used to classify whether an email is a phishing attempt:

**Prompt:**

```
Classify the following email to determine if it is a phishing attempt. Only respond with the exact words "Phishing" or "Not phishing."
"Dear Customer, Your account has been flagged for suspicious activity. Please verify your identity by clicking the secure link: https://bit.ly/secure-verification."
```

All of our models, including EUS Flash, can accurately classify this email as "Phishing." EUS Flash provides the correct classification on par with larger models, making it the most efficient and affordable option for tasks of this nature.

#### Specialized Industry Models: Tailored for Intermediate, Domain-Specific Tasks

UltraSafe AI's specialized industry models are perfect for handling intermediate tasks that demand both moderate reasoning and domain-specific expertise. Whether it's data extraction, document summarization, crafting job descriptions, or creating product descriptions, these models offer an optimal blend of performance and specialized knowledge. They are well-suited for tasks that require precise language transformation within specific industries.

For example, UltraSafe AI - Conversational model excels at generating personalized and contextually aware communications. For instance, it can be used to create a response to a customer inquiry:

**Prompt:**

<div align="left" data-full-width="false"><figure><img src="/files/p6OacQUA1UFsNZ3UPxCI" alt=""><figcaption></figcaption></figure></div>

**Output:**

<div align="left"><figure><img src="/files/Lwa7rhEmrDlMy0bqq9BT" alt=""><figcaption></figcaption></figure></div>

UltraSafe AI - Conversational is designed to generate responses that are not only accurate but also empathetic and tailored to the customer's needs, making it an excellent choice for customer service and engagement tasks.

#### EUS1: Tackling Complex, High-Level Reasoning and Specialized Tasks

EUS1 is the go-to model for handling intricate tasks that require substantial reasoning capabilities or highly specialized knowledge—such as advanced data analysis, strategic decision-making, or generating detailed reports.

**Complex Reasoning**

EUS1 excels in performing complex reasoning tasks, consistently outperforming other models in benchmarks for commonsense reasoning and logical analysis. Here’s an example where EUS1 demonstrates its ability to handle a multi-step reasoning process:

**Prompt:**

<div align="left"><figure><img src="/files/StizU7L0RtpBWU53X7Ao" alt=""><figcaption></figcaption></figure></div>

**Output:**

<div align="left"><figure><img src="/files/73KR59B67YI1f5QdB4Xj" alt=""><figcaption></figcaption></figure></div>

EUS1 is also capable of handling complex logical reasoning tasks. Here's another example:

**Prompt:**

<div align="left"><figure><img src="/files/iRrr903umYGeC4UEzWII" alt=""><figcaption></figcaption></figure></div>

**Output:**

<div align="left"><figure><img src="/files/0GT3gHFEWRjSHLjzsDqQ" alt=""><figcaption></figcaption></figure></div>

EUS1 also excels at complex logical reasoning tasks. Here’s another example:

**Prompt:**

<div align="left"><figure><img src="/files/l3oqVNEXiWxBvBamRhKR" alt=""><figcaption></figcaption></figure></div>

**Output:**

<div align="left"><figure><img src="/files/togoGidVZqLYneqac1pd" alt=""><figcaption></figcaption></figure></div>

**Conclusion**: Based on the given clues, the guilty suspect is Casey.

EUS1 is designed to handle such complex reasoning and problem-solving tasks efficiently, making it an indispensable tool for scenarios that require detailed analysis and strategic thinking.

**Coding**

UltraSafe AI - Code, renowned for its exceptional coding prowess, is the go-to choice for developers seeking powerful code generation and problem-solving capabilities in their projects.

Here's an example of UltraSafe AI - Code generating precise Python code based on a given prompt:

**Prompt:**

<div align="left"><figure><img src="/files/fqjEmSuGel9LLub42iJ3" alt=""><figcaption></figcaption></figure></div>

**Output:**

<div align="left"><figure><img src="/files/4jBSV7ijRCEgVPofleGQ" alt=""><figcaption></figcaption></figure></div>

Here's an example of UltraSafe AI - Code generating a function to compute square roots using the Babylonian method.

**Prompt:**

<div align="left"><figure><img src="/files/uM8HTVNTi2c9pDD66Fdo" alt=""><figcaption></figcaption></figure></div>

**Output:**

<div align="left"><figure><img src="/files/20pgp3qhBuCylDPwymwU" alt=""><figcaption></figcaption></figure></div>

#### Multi-lingual Tasks

In addition to its coding excellence, UltraSafe AI - Conversational demonstrates remarkable proficiency in handling multilingual tasks. Trained to comprehend and generate text in over 100 human languages, it serves as an invaluable asset for businesses and users operating across diverse linguistic landscapes.

**Prompt:** Translate the following English sentence into Spanish, French, and German: "The quick brown fox jumps over the lazy dog."

**Output:** Spanish: "El rápido zorro marrón salta sobre el perro perezoso." French: "Le rapide renard brun saute par-dessus le chien paresseux." German: "Der schnelle braune Fuchs springt über den faulen Hund."

UltraSafe AI - Conversational not only delivers accurate translations but also maintains contextual appropriateness, ensuring effective communication across multiple languages. This capability makes it an essential tool for global businesses, international organizations, and multilingual content creators.