NovatasticRoScript/Atomight-V2.1-0.5B-Inference
NovatasticRoScript's Atomight-V2.1-0.5B-Inference is an ultra-compact, 494M parameter causal language model derived from Qwen, optimized for reasoning tasks. Refined with GRPO reinforcement tuning, it targets highly efficient edge-device reasoning, structured text outputs, and lightweight coding assistance. This model excels in environments with severe compute constraints, demonstrating strong performance in scientific fact retrieval and commonsense reasoning.
Loading preview...
Atomight-V2.1-0.5B-Inference: Ultra-Compact Reasoning Model
Atomight-V2.1-0.5B-Inference is a 494 million parameter causal language model developed by NovatasticRoScript, part of the Atomight Ecosystem. Built on a Qwen-derived foundation, it has been specifically refined using GRPO (Group Relative Policy Optimization) reinforcement tuning to prioritize high-signal reasoning vectors over brute-force dataset scale. This design makes it exceptionally efficient for edge-device deployment.
Key Capabilities & Differentiators
- Ultra-Compact & Edge-Optimized: With ~494M parameters, it loads into approximately 1GB VRAM (FP16), making it ideal for low-overhead mobile, local, and browser-based inference.
- Reasoning-Oriented: The model's GRPO training focuses on enhancing reasoning capabilities, structured text outputs, and lightweight coding assistance.
- Competitive Benchmarking: Despite its small size, Atomight-V2.1-0.5B-Inference demonstrates strong performance in reasoning tasks. It achieves 59.34% on ARC-Easy and 33.79% on ARC-Challenge. Notably, it surpasses Meta's larger Llama-3.2-1B-Instruct on localized logic-retrieval metrics (ARC-Easy and ARC-Challenge).
- Mathematical Accuracy: On GSM8K with flexible extraction, it scores 32.45%, outperforming both Qwen2.5-0.5B-Instruct (26.8%) and Llama-3.2-1B-Instruct (24.4%) in raw mathematical accuracy, even if its strict formatting score is lower due to dense internal thinking traces.
Ideal Use Cases
- Resource-Constrained Environments: Perfect for applications requiring powerful reasoning on devices with limited computational resources.
- Structured Output Generation: Suited for tasks where structured text responses are beneficial.
- Lightweight Code Assistance: Can provide efficient coding support without the overhead of larger models.