> ## Documentation Index
> Fetch the complete documentation index at: https://student-213fb9fc.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Model Families

> Understanding the AI model families available on AJ STUDIOZ Cloud Infra

## Model Families Overview

AJ STUDIOZ Cloud Infra hosts models from six major AI research labs, each with distinct strengths.

<CardGroup cols={3}>
  <Card title="Gemma (Google)" icon="google" color="#4285F4">
    Lightweight, efficient models from Google DeepMind. Great for production workloads.
  </Card>

  <Card title="Qwen (Alibaba)" icon="microchip-ai" color="#FF6B00">
    High-capability coding and multilingual models with massive scale.
  </Card>

  <Card title="Kimi (Moonshot)" icon="moon" color="#1A1A2E">
    Ultra-large models excelling at agentic and reasoning tasks.
  </Card>

  <Card title="DeepSeek" icon="magnifying-glass" color="#00A8E8">
    Top-tier open-source models for coding and analysis.
  </Card>

  <Card title="GLM (Zhipu AI)" icon="brain" color="#7C3AED">
    Multilingual long-context models with strong Chinese language support.
  </Card>

  <Card title="Mistral & Others" icon="wind" color="#FF4500">
    Best-in-class European AI models with multilingual excellence.
  </Card>
</CardGroup>

***

## Gemma Family (Google DeepMind)

Gemma models are open-weight, production-ready models from Google. They offer excellent quality-per-compute-cost ratio.

| Model        | Parameters | Notes                                           |
| ------------ | ---------- | ----------------------------------------------- |
| `gemma3:4b`  | 4B         | Fastest Gemma, ideal for real-time applications |
| `gemma3:12b` | 12B        | Balanced performance                            |
| `gemma3:27b` | 27B        | Best Gemma quality, recommended default         |

**Best for:** General assistant tasks, fast inference, cost-sensitive workloads.

***

## Qwen Family (Alibaba Cloud)

Qwen models offer state-of-the-art performance especially in coding and mathematical reasoning.

| Model                    | Parameters | Notes                          |
| ------------------------ | ---------- | ------------------------------ |
| `qwen3-coder-next`       | \~82B      | Next-gen coding model          |
| `qwen3-coder:480b`       | 480B       | Largest coding model available |
| `qwen3-next:80b`         | 80B        | Balanced coding + reasoning    |
| `qwen3.5:397b`           | 397B       | Latest Qwen release            |
| `qwen3-vl:235b`          | 235B       | Vision + language              |
| `qwen3-vl:235b-instruct` | 235B       | Vision, instruction-tuned      |

**Best for:** Code generation, debugging, mathematical problems, vision tasks.

***

## Kimi Family (Moonshot AI)

Kimi models are among the largest available, optimized for complex multi-step reasoning and agentic workflows.

| Model              | Parameters | Notes                                |
| ------------------ | ---------- | ------------------------------------ |
| `kimi-k2:1t`       | \~1.1T     | Largest model, general agentic tasks |
| `kimi-k2.5`        | \~1.1T     | Enhanced capabilities                |
| `kimi-k2-thinking` | \~1.1T     | Optimized for chain-of-thought       |

**Best for:** Complex reasoning, research, multi-step agentic tasks, long context.

***

## DeepSeek Family

DeepSeek produces highly capable open-source models that compete with GPT-4 class models.

| Model                | Parameters | Notes                             |
| -------------------- | ---------- | --------------------------------- |
| `deepseek-v3.1:671b` | 671B       | High-quality coding and reasoning |
| `deepseek-v3.2`      | 671B       | Latest DeepSeek release           |

**Best for:** Coding, data analysis, structured outputs, STEM tasks.

***

## GLM Family (Zhipu AI)

GLM models excel at long-context understanding and multilingual tasks, especially Chinese.

| Model     | Parameters | Notes                         |
| --------- | ---------- | ----------------------------- |
| `glm-5`   | Very Large | Latest flagship, long context |
| `glm-4.6` | Large      | Mature, stable release        |
| `glm-4.7` | Large      | Enhanced version              |

**Best for:** Long documents, Chinese language, multilingual tasks.

***

## MiniMax Family

MiniMax models offer strong multimodal capabilities.

| Model          | Parameters | Notes                  |
| -------------- | ---------- | ---------------------- |
| `minimax-m2`   | 230B       | Multimodal baseline    |
| `minimax-m2.1` | 230B       | Enhanced multimodal    |
| `minimax-m2.5` | 230B       | Latest MiniMax release |

**Best for:** Vision + language tasks, multimodal reasoning.

***

## Other Models

| Model                    | Family     | Notes                          |
| ------------------------ | ---------- | ------------------------------ |
| `gpt-oss:120b`           | OpenAI OSS | Open-source OpenAI-style model |
| `gpt-oss:20b`            | OpenAI OSS | Efficient variant              |
| `cogito-2.1:671b`        | Cogito     | Scientific reasoning           |
| `devstral-2:123b`        | Mistral    | Coding-specialized             |
| `devstral-small-2:24b`   | Mistral    | Efficient coding model         |
| `mistral-large-3:675b`   | Mistral    | Flagship multilingual          |
| `ministral-3:3b`         | Mistral    | Ultra-compact                  |
| `ministral-3:8b`         | Mistral    | Compact balanced               |
| `ministral-3:14b`        | Mistral    | Efficient mid-size             |
| `nemotron-3-nano:30b`    | NVIDIA     | Optimized inference            |
| `rnj-1:8b`               | Community  | Compact reasoning              |
| `gemini-3-flash-preview` | Google     | Ultra-fast API model           |

***

## Choosing the Right Model

<Accordion title="For chat and general tasks">
  **Start with:** `gemma3:27b`

  Great balance of quality and speed. If you need more capability, try `deepseek-v3.2` or `glm-5`.
</Accordion>

<Accordion title="For coding tasks">
  **Start with:** `qwen3-coder:480b` for best quality, or `devstral-2:123b` for a balance of speed and quality.

  For small, fast coding tasks: `devstral-small-2:24b` or `gemma3:12b`.
</Accordion>

<Accordion title="For agentic / research tasks">
  **Use:** `kimi-k2:1t` or `kimi-k2-thinking` for complex multi-step reasoning.

  Alternatively, `cogito-2.1:671b` for scientific domains.
</Accordion>

<Accordion title="For fast/cheap inference">
  **Use:** `gemma3:4b`, `ministral-3:3b`, or `gemini-3-flash-preview`.

  These models prioritize speed and cost over raw capability.
</Accordion>

<Accordion title="For vision tasks">
  **Use:** `qwen3-vl:235b-instruct` or `minimax-m2.5`.
</Accordion>