README.md

---
title: README
emoji: 🏃
colorFrom: purple
colorTo: yellow
sdk: static
pinned: false
---

# NitrAI ⚡

Welcome to **NitrAI**. We are an AI research and development organization dedicated to bridging the gap between frontier closed-source intelligence and consumer-grade hardware. 

Our core focus is **efficient reasoning distillation** — capturing complex coding-agent trajectories, multi-step mathematical logic, and system-level reasoning from massive frontier LLMs and packing them into highly optimized, lightweight models you can actually run locally.

---

## 🎯 Our Mission

*   **Frontier Distillation:** We extract high-level cognitive patterns from state-of-the-art models (such as GPT-5.5, Claude-Fable-5, and GLM-5.2) into accessible open-weights architectures.
*   **Consumer-First Optimization:** High-fidelity reasoning shouldn't require data-center-scale infrastructure. We build for local execution on everyday GPUs and CPUs.
*   **Agentic & Math Focus:** Our models are specifically fine-tuned for agentic workflows, long-context system analysis, and rigorous logical reasoning.

---

## 🚀 Featured Models

### 🧠 OpenGCM-v2 (9B)
**OpenGCM-v2** is our flagship reasoning-focused model. It is engineered to deliver enterprise-grade logical reasoning and coding proficiency within a lightweight 9-billion parameter budget.

*   **Base Architecture:** Qwen3.5-9B
*   **Context Window:** 262k tokens (ideal for analyzing large codebases and complex system logs)
*   **Core Capabilities:** 
    *   Distilled multi-step math logic and proof generation.
    *   Complex coding-agent trajectories and system-level debugging.
    *   High-efficiency inference on consumer hardware.

### 🌟 Polaris-V1 (4B)
**Polaris-V1** is designed to redefine the boundaries of lightweight local intelligence. Engineered to deliver near-frontier capabilities within a highly efficient 4-billion parameter envelope, it bridges the gap between extreme context length and uncompromising reasoning quality.

*   **Base Architecture:** Qwen3.5-4B
*   **Context Window:** 1,592,638 tokens (1.5M+ context utilizing precision-focused YaRN-scaling)
*   **Core Capabilities:**
    *   **Extreme-Scale Retrieval:** Flawless "Needle in a Haystack" performance across millions of tokens, making it capable of analyzing entire multi-repo codebases in a single prompt.
    *   **Premium 2026 Distillation:** Fine-tuned on a state-of-the-art dataset distilled from elite frontier models (including Kimi K3, Qwen3.8), bypassing outdated GPT-3.5/4 patterns entirely.
    *   **Interactive ChatML Workflows:** Fully conversational agentic reasoning, moving beyond simple text completion into highly precise, multi-turn system debugging and instruction-following.
    *   **Hardware-Optimized Local Run:** Specifically tailored for lightning-fast local inference on consumer GPUs using optimized custom kernels and native bfloat16 execution.

---

## 🛠️ Getting Started

### Quick Inference with Hugging Face & Transformers
You can easily load our models using the `transformers` library:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "NitrAI/OpenGCM-v2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name, 
    device_map="auto", 
    torch_dtype="auto"
)

prompt = "Analyze the following system trace and identify the deadlock:"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))