NorviTech

Open-weight local tool-calling models, sub-millisecond neural routers, and sealed template-disjoint benchmarks for Arrmes (v1.7.2 / v10).

• NorviTech/arrmes-neural-router-v3 — 7.48 MiB temperature-calibrated CPU multi-head hashed n-gram MLP + structural verifier (97.86% verified positive and 96.14% verified hard-negative accuracy on sealed acceptance-v8 N=3,135; 1,380/1,380 positive and 1,200/1,200 v3 hard-negative accuracy on acceptance-v7 N=2,580 plus 80/80 on overnight-v1 with three-platform bitwise-identical outputs across macOS ARM64, Linux x86_64, and Windows Server 2025 AMD64; 0.416 ms M4 Pro CPU p50, 0.714 ms Ryzen 9 7950X3D CPU p50, 79.27 MiB Win32 peak working set).
• NorviTech/arrmes-qwen3-0.6b — Unassisted Qwen3-0.6B v10 28-class BF16 LoRA SFT + NLL-anchored RPO-DPO (mlx-4bit/, Q8_0, and Q4_K_M GGUF; 84.21% verified positive and 95.24% verified hard-negative accuracy on sealed acceptance-v8 with 0 unauthorized verified writes; 99.33% verified hard-negative accuracy on acceptance-v7; 97.50% verified on overnight-v1 BF16/MLX-8bit/Q8_0 and 88.75%–95.00% across macOS, Linux CPU/CUDA, and native Windows x64 Q4_K_M).
• NorviTech/arrmes-functiongemma-270m — Unassisted FunctionGemma-270M v10 28-class BF16 LoRA SFT + positive-anchored RPO-DPO (mlx-4bit/, Q8_0, and Q4_K_M GGUF; 94.08% verified hard-negative accuracy on acceptance-v7 with 0 unauthorized writes; 88.75%–90.00% verified on overnight-v1 Q4_K_M across macOS, Linux CPU 131.32 ms p50, CUDA 40.73 ms p50, and native Windows x64 CPU 539.29 MiB Win32 peak working set).
• datasets/NorviTech/arrmes-bench-v8 & datasets/NorviTech/arrmes-bench-v7 — Publicly accessible template-skeleton-disjoint evaluation suites (N=3,135 sealed acceptance-v8 and N=2,580 acceptance-v7 + N=80 overnight-v1).