English
Guides
Six notes on fitting a model to a machine. Memory numbers are the formula. Prices are indicative and unsourced. Nothing here is a measured benchmark.
- Best GPU for local LLMs in 2026
Which GPU in the localLLMrig catalog can hold an 8B, 32B or 70B model in 2026. Fit comes from the formula, not from a measured benchmark.
- How much VRAM does a 70B model need?
Weight and KV-cache estimates for Llama 3.3 70B, Qwen2.5 72B and the DeepSeek 70B distill at Q4, Q5, Q8 and FP16. No measured benchmark.
- Mac vs PC for a local LLM
Unified memory against discrete VRAM, using Apple spec pages where we stored them and the localLLMrig fit formula. Speed stays an estimate.
- Used RTX 3090 buying guide (EU)
What a used 24 GB RTX 3090 does and does not hold in this catalog, and the indicative EU prices dated 2026-10-14. No shop listing was checked.
- Running DeepSeek and Qwen locally
Which DeepSeek and Qwen models are actually in the localLLMrig catalog, which configs were fetched, and what memory the formula asks for.
- Quantization explained for local LLMs
What Q4, Q5, Q8 and FP16 mean in the localLLMrig formula, with weight estimates for an 8B and a 70B model. Quality loss is not measured here.
- The RTX 5090 is (reportedly) dead: best 32GB alternatives for local LLMs
A 10 October 2026 leak says NVIDIA is ending GeForce RTX 5090 production. NVIDIA has not confirmed it. The 32 GB alternative with a vendor spec page is the Radeon AI PRO R9700. A used RTX 3090 is 24 GB, not 32.
- Is RTX Spark worth it for local LLMs vs Mac Studio vs Strix Halo?
NVIDIA’s RTX Spark page says up to 128 GB of unified memory and up to 1 petaflop of FP4. No laptop price was on that page. Mac Studio rows in the catalog and AMD’s Ryzen AI Max+ 395 spec are cited separately.
- Run Nemotron 3.5 Bolt 74B on a 24GB GPU
NVIDIA says Nemotron 3.5 Bolt 74B-A9B arrives 15 October 2026 in a 24 GB 2-bit pack. A Hugging Face search on 14 October 2026 found no Bolt weights, so this page does not pretend you can download it yet.
- Run a local LLM on Windows in 5 minutes (Llama Windows)
Hugging Face’s Llama Windows article, dated 13 October 2026, is a tray app for llama.cpp on Windows 11. The article does not time the steps, so five minutes is not a measured claim.
- Best local LLM rig under $500 / $1000 / $3000
Catalog rows under those dollar caps, converted with the ECB rate of 14 October 2026. Euro prices are indicative and have no shop source. Nothing here is a measured ranking.