jcbtc/Step-3.7-Flash-ROCmFPX-Q3-QualityPlus

1{2  "model": "Step-3.7-Flash-ROCmFPX-Q3-QualityPlus",3  "benchmark": "Tool-Eval full",4  "harness_version": "2.0.7",5  "run_id": "2026-07-05T06-07-28.330900Z_405f5a32",6  "scenario_count": 69,7  "final_score": 88,8  "deployability": 70,9  "responsiveness": 27,10  "total_points": 122,11  "max_points": 138,12  "status_counts": {13    "pass": 57,14    "partial": 8,15    "fail": 416  },17  "runtime": {18    "backend": "llama.cpp ROCmFPX / Vulkan",19    "context": 65536,20    "mtp": true,21    "speculative_n_max": 2,22    "speculative_p_min": 0.75,23    "target_kv": "q8_0/q8_0",24    "draft_kv": "q8_0/q8_0",25    "batch": 8192,26    "ubatch": 204827  },28  "speed_and_memory": {29    "pp_tps": 145.65656394243823,30    "tg_tps": 32.338072639262656,31    "prompt_tokens": 34333,32    "predicted_tokens": 33430,33    "peak_pooled_gpu_gib": 96.34914779663086,34    "peak_ram_used_gib": 107.00336074829102,35    "wall_time": "21:21.23"36  },37  "notes": [38    "Local AMD Ryzen AI Max+ 395 / Strix Halo measurement.",39    "Uses the Step native tool_response chat template with protocol-boundary escaping.",40    "The public llm.ciru.ai StepFun tool-eval page currently documents the Step tool-calling methodology and matching FP4 score row; this JSON records the exact Q3 QualityPlus run summary."41  ]42}43