Evaluation Setup: All models evaluated on identical hardware (NVIDIA DGX Spark GB10, concurrency 32, zero cache hits) using official verifiers across standardized benchmark questions. Primary baseline: nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16 (14.9 GB unquantized reference).
Evaluated with full image and raw audio waveform inputs against unquantized Google Gemma 4 E4B-IT BF16 reference.
Hardware Target: NVIDIA GeForce RTX 5090 Blackwell Workstation (32 GB GDDR7 @ 1,792 GB/s, SM_120).
Evaluation Mode: Pure text inference without multimodal projector (--mmproj omitted), ensuring 100% fair baseline with zero vision overhead.
Strict Metric Separation: Prompt Processing (PP / prefill) and Token Generation (TG / decode) are strictly separated; they are never combined into a blended metric.
Hardware Target: NVIDIA DGX Spark Grace Blackwell GB10 Datacenter Node (128 GB LPDDR5x unified memory @ 273 GB/s, SM_121a).
All models evaluated natively on cdemy-spark (192.168.1.110) in Pure Text Mode:
This model is published under the Open Model Derivative License (OpenMDW-1.1), incorporating permissions and attribution from upstream NVIDIA Open Model License Agreement and Google Gemma Apache 2.0 license.