nitrai-research/OpenGCM-v2
OpenGCM-v2 is a 9 billion parameter reasoning-focused model developed by NitrAI, built upon the Qwen3.5-9B base model with a 262k context window. It specializes in distilling complex coding-agent trajectories, multi-step mathematical logic, and system-level reasoning from frontier LLMs. This model is optimized for efficient performance on consumer hardware, excelling in math reasoning and localized code bug verification.
Loading preview...
Overview
OpenGCM-v2 is a 9 billion parameter model from NitrAI, based on the Qwen3.5-9B architecture, designed for advanced reasoning tasks. It was created by distilling complex reasoning patterns, coding-agent trajectories, and multi-step math logic from powerful frontier LLMs like GPT-5.5, Claude-Fable-5, and GLM-5.2. The goal is to provide these capabilities in a lightweight, consumer-hardware-friendly package.
Key Features & Training
- Base Model: Qwen3.5-9B with a 262k context window.
- Distillation Dataset: Fine-tuned on a highly curated dataset of 597 high-signal QA items (904,466 tokens) focusing on agent trajectories, detailed reasoning, and complex system-level traces.
- Training Methodology: Utilizes Unsloth library and DoRA (Weight-Decomposed Low-Rank Adaptation) for efficient training on consumer GPUs.
Performance & Strengths
OpenGCM-v2 demonstrates strong performance in specific reasoning benchmarks:
- Exceptional Math Reasoning: Achieved 100% accuracy on AIME 26, solving sequence problems perfectly.
- Code Reasoning: Highly capable in localized code reasoning and bug patch verification, scoring 100% on SWE-bench Pro.
Limitations
- May exhibit occasional instability or context drift during extremely long inference generations. NitrAI recommends using a lower temperature (e.g., 0.2 or 0.4) and structured system prompts to mitigate this.