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.

Gemma (Google)

Lightweight, efficient models from Google DeepMind. Great for production workloads.

Qwen (Alibaba)

High-capability coding and multilingual models with massive scale.

Kimi (Moonshot)

Ultra-large models excelling at agentic and reasoning tasks.

DeepSeek

Top-tier open-source models for coding and analysis.

GLM (Zhipu AI)

Multilingual long-context models with strong Chinese language support.

Mistral & Others

Best-in-class European AI models with multilingual excellence.

---

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

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.

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.

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.

For fast/cheap inference

Use: gemma3:4b, ministral-3:3b, or gemini-3-flash-preview.

These models prioritize speed and cost over raw capability.

For vision tasks

Use: qwen3-vl:235b-instruct or minimax-m2.5.