The new BMT LLM Reference technical guide has been released, classifying 38 modern large language models by deployment type, VRAM requirements, and specializations.
What Happened
BMT LLM Reference provides a detailed classification of 38 current LLMs, dividing them into cloud-based and local solutions. The document contains specific hardware selection recommendations: for example, 7–12B class models (such as Qwen 3 7B or Llama 8B) require 8–12 GB of VRAM, while running top-tier local models like Llama 4 Maverick requires a multi-GPU configuration with 96 GB of memory or more. The guide also highlights leaders among cloud services, including GPT-5.5, Claude Opus 4.8, and Gemini 3.1 Pro.
Context
With the growing diversity of models and deployment methods, there has been a need for a standardized way to assess the alignment of LLM architectural capabilities with available hardware resources. This allows for a transition from empirical guesswork to calculated parameters when designing AI infrastructure.
Why It Matters for the Industry
For the industry, the emergence of such a tool means creating a technical bridge between theoretical model power and practical hardware capabilities. This enables optimization of inference costs and accelerates decision-making when choosing between using APIs and self-hosted solutions. In the long term, such data could become a 'hardware-model mapping' standard and be integrated into orchestration systems like Kubernetes or KServe.
Why It Matters for Users
Developers and engineers receive a ready-made guide for quickly calculating costs and hardware configurations for specific tasks. The tool allows for an instant assessment of the technical feasibility of running heavy models on existing hardware or selecting the optimal GPU stack for new AI products—ranging from lightweight models like Phi-4 Mini on mobile devices to powerful local systems.
What Is Not Yet Known / Limitations
There is a noticeable difference in focus during data analysis: while technical specialists are oriented toward VRAM and architectures, business roles and legal consultants are more focused on CapEx/OpEx issues and compliance risks when changing deployment types.
Sources
Author
Look at AI, Editorial Staff