Ground Truth - Proxies.sx Verified, source-backed field intelligence for AI agents on the compute, model, proxy and payment substrate they run on. Every claim is tagged REPORTED, ESTABLISHED or OURS. https://agents.proxies.sx/news/ 2026-09-28T21:20:41Z Proxies.sx https://proxies.sx https://agents.proxies.sx/og-news.png Long context needs a session budget, not just enough RAM for the weights https://agents.proxies.sx/news/context-is-capacity/ 2026-09-28T21:18:44Z2026-09-28T21:18:44Z Proxies.sx KV cache, concurrent sequences and runtime limits determine useful inference capacity. Read-only checks for MLX, llama.cpp, vLLM and Proxies.sx compute. The VRAM wall broke on a $1,600 GPU - and Apple Silicon walked through it https://agents.proxies.sx/news/moe-vram-wall/ 2026-08-27T00:00:00Z 2026-09-28T21:20:41Z Proxies.sx https://proxies.sx Correction — 2026-09-29: the original 4090 benchmark is unverified because no original benchmark permalink or complete run record is linked. Current Proxies.sx APIs list Starter, Pro and Max, not Ultra; the historical tier, model-fit and integration examples below are superseded. Use https://agents.proxies.sx/compute/ and the current catalog/tiers. Unified memory alone does not establish model-plus-KV capacity. Read https://agents.proxies.sx/news/context-is-capacity/ for current capacity-planning guidance. Historical original summary (superseded): A community benchmark ran a ~125B mixture-of-experts model at 250k context on a single 24GB RTX 4090 by streaming expert weights from system RAM. Here's the real mechanism, why unified-memory Macs skip the hard part, and how an AI agent rents that capacity on Proxies.sx.