/blog/research
Technical thinking, architecture notes, and operating lessons from production AI and applied systems work.
ResearchA practical benchmark comparing a local NVIDIA GB10 AI agent running Qwen3.6 35B through Hermes against GPT-5.5 on real office workflows, including code repair, document synthesis, web edits, travel research, and concurrency limits.
ResearchA practical RTX 4090 vs NVIDIA GB10 local AI benchmark comparing Qwen3.6 35B Q4, Q5, MXFP4, long-context workloads, and dense 70B inference. The results show where the RTX 4090 dominates, where GB10’s memory headroom matters, and why the VRAM cliff is one of the most important limits in local AI hardware.
ResearchA detailed benchmark of Qwen3.6-35B-A3B on NVIDIA GB10 using 243 llama.cpp runs, comparing BF16, Q8, Q5, Q4, and MXFP4 GGUF variants for local AI inference, RAG, agents, and long-context workloads.
ResearchA hands-on benchmark of Gemma 4 on NVIDIA GB10 comparing 31B dense and 26B-A4B MoE variants across speed, memory, thinking mode, and practical deployment tradeoffs.
A deep dive into how we architect autonomous AI agent systems for enterprise clients — from task decomposition to fault-tolerant execution.
As AI adoption accelerates, organizations need governance frameworks that enable innovation while managing risk. Here's our approach.
Subterra AI Ask anything about our services, tech, or how we can help.