Modelpublic

AI-ku-official/AI-ku_V4X_Thinker_Max_35B

sourceHugging Faceupdated 12d agoView on Hugging Face

0likes

Model Card

AI-ku V4X Thinker Max

<a href="#1-introduction">Introduction</a> • <a href="#2-model-summary">Model Summary</a> • <a href="#3-ecosystem--plug--work">Ecosystem</a> • <a href="#4-evaluation-results">Evaluation</a> • <a href="#5-how-to-run-locally">How to Run Locally</a>

1. Introduction

We introduce AI-ku V4X Thinker Max, a 35B-parameter MoE model designed from the ground up as a multimodal agentic system with advanced reasoning and task-execution capabilities.

While recent reasoning models have demonstrated powerful Chain-of-Thought (CoT) capabilities, they often suffer from severe context bloat and memory exhaustion during long tasks. To address this, AI-ku V4X introduces the ACS (AI-ku Context Saver), a revolutionary architectural approach that truncates historical reasoning steps in the backend, providing almost limitless effective context efficiency within its massive 1M token context window.

Combined with dynamic cognitive scaling and out-of-the-box native tool calling, AI-ku V4X bridges the gap between deep analytical reasoning and seamless agentic deployment.

2. Model Summary

Cognitive Architecture: Dynamic Reasoning

AI-ku V4X does not treat all prompts equally. It employs a dynamic computing approach to maximize efficiency:

—ToD (Thinking on Demand): a dynamic reasoning policy that determines when additional reasoning effort is warranted.

Memory & Context Efficiency: The ACS System

Long-running agentic tasks (10h+) typically collapse under the weight of their own reasoning tokens. The AI-ku Context Saver (ACS) solves this. By securely pruning obsolete internal monologues from past turns while preserving the final visible outputs and the 1M context length, ACS allows the model to think as deeply as necessary without cannibalizing its future memory.

Multimodal & Agentic Capabilities

—Native Tool Calling: Supports structured invocation of external APIs, local scripts, and system-level tools.
—Multimodal Ready: Deeply integrated to support vision (image and video analysis) and audio/speech recognition (interfacing with tools like Mert and Whisper).

3. Ecosystem & Plug & Work

AI-ku V4X is not just a weight checkpoint; it is a complete, deployable agentic ecosystem available at Tito-Gl1tch3r's GitHub (Links TBD).

—Plug & Work Python Suite: Comes out of the box with ready-to-use Python scripts.
—Telegram Integration: Deploy your autonomous agent directly to Telegram with zero friction.
—Media Generation Hooks: Pre-configured to seamlessly trigger image and media generation pipelines based on conversational context.

4. Evaluation Results

Benchmarks and rigorous evaluation metrics are actively being compiled by the research team. Stay tuned for comprehensive results comparing AI-ku V4X against state-of-the-art dense and reasoning models.

(Note: [Insert Gl1tch3r's Benchmarks Here])

5. How to Run Locally & Usage Recommendations

AI-ku V4X features a 35B-parameter MoE architecture with 35B active parameters, optimized for accessible local hardware deployment.

Deployment Options

—AirLLM (Highly Recommended for Low VRAM): For users with constrained GPU memory, we strongly recommend deploying AI-ku via AirLLM. This ensures stable, layer-wise inference without OOM (Out of Memory) crashes.
—Ollama & Standard Inference: Fully compatible with Ollama and standard inference engines for environments with sufficient compute.

Usage Recommendations

To achieve the expected agentic performance, please adhere to the following guidelines:

1.Trust the ACS: Do not attempt to forcefully limit the model's output length for complex tasks. Let the model reason freely; the ACS will handle the historical context economy.
2.Do Not Manually Prompt Reasoning Tokens: The model natively handles its own start and end reasoning syntax. Do not inject custom thought tags into your system prompts; allow the ToD engine to route cognition naturally.
3.Temperature Settings: For standard coding and reasoning tasks, a lower temperature (e.g., 0.3 - 0.5) is recommended to maintain logical consistency.

6. License & Contact

Developed by Gl1tch3r (Tito-Gl1tch3r).