model-zoo

---
  Date: 04-26-2026
  Version: 1.0
---

CreativeAct Model Zoo

Manifesto: A curated repository of high-performance LLMs, specialized for local deployment, agentic workflows, and modular engineering. This index prioritizes open-weight models while providing proprietary benchmarks for comparative analysis.

Purpose	GitHub Repository	Last Updated	Total Models
Open-Source AI Resources	CreativeAct/model-zoo	04/26/2026	17

---

🛠 Open-Weight & Local Deployment Models

These models allow for weight inspection, quantization, and local hosting (Ollama, vLLM, Llama.cpp).

1. GLM-5.1 (General Language Model)

• Architecture: Sparse Mixture-of-Experts (MoE)
• Parameters: 744B
• Context Window: 128k+
• Agentic Specialty: Yes (Native generation of structured PDFs/Spreadsheets)
• Highlight: Uses the "Slime" asynchronous RL framework to minimize hallucination rates in complex system management.

2. NVIDIA Nemotron

• Architecture: Optimized for Blackwell/H200 (NVFP4 Format)
• Primary Use: Industrial RAG, Synthetic Data Generation, Multi-Agent Systems.
• Agentic Specialty: Yes
• Technical Edge: Engineered for extreme throughput in enterprise-grade autonomous environments.

3. Mistral / Mixtral

• Architecture: Sparse MoE
• Context Window: 256k
• Agentic Specialty: Yes (Model Context Protocol / MCP support)
• Primary Use: Production-grade RAG and tool-calling efficiency.

4. Qwen (Alibaba)

• Context Window: 1,000,000 tokens
• Primary Use: Multimodal data extraction, long-form code analysis, and complex reasoning.
• Agentic Specialty: High (Strong performance in function-calling benchmarks).

5. IBM Granite

• Primary Use: Enterprise workflow automation, regulatory compliance, and governance.
• Highlight: Optimized for dense retrieval and "safe" enterprise deployments.

6. DeepSeek

• Primary Use: Agentic coding and mathematical reasoning.
• Highlight: Highly efficient training-to-performance ratio; preferred for logic-heavy autonomous tasks.

7. Liquid AI (Leap Workbench)

• Architecture: Liquid Neural Networks (Non-Transformer)
• Highlight: Constant-time complexity; highly efficient for edge deployment and time-series data analysis.

8. Meta (Llama Series)

• Primary Use: Foundation for fine-tuning and specialized domain adaptation.
• Highlight: The industry standard for open-weight ecosystem compatibility.

9. StepFun (Step-3.5-Flash / Step-2)

• Primary Use: Real-time agentic problem solving and multimodal vision.
• Highlight: Focuses on "Intelligence Density" to provide trillion-parameter reasoning speeds in smaller footprints.

---

🏗 Hubs & Managed Endpoints

Orchestration layers for multi-model inference and testing.

• OpenRouter.ai: Unified API for accessing hundreds of open and closed models.
• Google AI Studio: Developer sandbox for Gemini-class models and multimodal experimentation.
• Liquid AI Playground: Testing environment for non-transformer architectures.

---

📊 Proprietary Benchmarks (Reference Only)

SaaS-based models for performance validation and comparative testing.

Model	Provider	Primary Use-Case
Claude	Anthropic	Coding, nuanced reasoning, high-fidelity instruction following.
ChatGPT	OpenAI	General-purpose assistance, rapid prototyping.
Gemini	Google	Real-time research, massive context (up to 2M), multimodal native.
Perplexity	Perplexity	Real-time web-indexed search and citation-heavy research.
Ernie 5	Baidu	Large-scale knowledge retrieval (Bilingual focus).
Kimi K2.6	Moonshot	Long-context memory and Chinese language processing.

---

🔧 Engineering Integration

To integrate these models into an autonomous workflow, refer to the following standardized specifications in the root directory:

• AGENTS.md - Standardized tool-calling schemas.
• SKILL.md - Portable skill definitions for browser-based agents.

© 2026, CreativeAct Technologies