LLM Playground - Inside the Architectures (Pt.10) — Silas Data

A benchmark study of twelve open-weight language models releasable on a single 8 GB GPU found that architecture choices now vary widely across dense, mixture-of-experts (MoE), Mamba-2, recurrent and ternary-weight designs, with parameter counts ranging from 1.7B to 35B. The lineup includes Llama 3.1 8B and 3.2 3B, Qwen2.5 7B, Qwen3-VL 4B and Qwen3.6-35B-A3B, Gemma 3 4B and Gemma 4 E4B, Phi-4-multimodal, gpt-oss-20b, Nemotron 3 Nano Omni, LFM3 1.7B and Ternary Bonsai 2 27B, spanning releases from July 2024 to October 2026. The study frames the trade-off in terms of weights and KV-cache memory, noting that dense models must read every parameter per token, so memory footprint and per-token cost scale with size. It matters because engineers selecting models for constrained hardware must weigh architecture efficiency against raw scale rather than parameter count alone. Next, the study is expected to report measured performance and resource costs for each architecture.