Download more VRAM. Issue 2’s cover story is the CMP 170HX, an 8 GB mining card that unlocks to 64 GB: which card you have, what fits on one, two and four of them, and every step of the setup.

Other calculators guess your context. This guide reads it from each lab’s config. A popular rule of thumb budgets 31.8 GB of cache for Qwen 3.8 27B at 64K; llama.cpp needs 4.2 GB, a guess 7.5 times too high.

Inside Issue 2:
Strata: a 180B model on an RTX 5090 with 128 GB of RAM, and every cell of our own Strata runs.
DeepSeek V4.1 Flash 1017 on every machine measured, from a 16 GB card to four RTX PRO 6000.
The GPU unlock: 82 to 155 BF16 TFLOPS on an RTX 3090, and what it does for a local LLM.
Old Xeons, DDR3 and two modded mining cards at 75 tok/s, and what such a box costs now.
The M5 Ultra against its rivals, priced from listings you can buy.
Every model on every tier from 8 GB to 128 GB and on eight unified-memory machines, at 32K of context.

Every model checked against its lab’s own config file.

Issue 2, 74 pages, and the 67-page September issue are in your Whop library as soon as you pay. The November issue is included: three issues for one purchase.