Kinetic-4B beats Claude Haiku 4.5 on tool calling

On a 300-sample Composio benchmark, the 4B-parameter Kinetic-4B from Bengaluru lab Conscious Engines scored 82.33% accuracy at 1.61s p95 latency, versus 80.0% and 4.02s for Claude Haiku 4.5 and 76.33% and 7.99s for GPT-OSS-120B. Training took about 4.5 hours on one rented GPU via LoRA, and the adapter is published openly on Hugging Face.

• Kinetic-4B: 82.33% accuracy, 4.67% failed calls, 1.61s p95
• Claude Haiku 4.5: 80.0% accuracy, 9.67% failed calls, 4.02s p95
• GPT-OSS-120B: 76.33% accuracy at 7.99s p95
• LoRA training: 132M parameters, ~4.5 hours on one GPU