Local LLM · GPU · Shops in Japan · 2026-09
The most reliable way to own an "uncensored" AI is to run the model on a PC in your own home. Nothing passes through a cloud provider, so there are no provider filters, no usage logs on someone else's server, and no per-token billing. What you need instead is VRAM. This guide covers three things: ① how much memory you need for which models, ② what to check when buying used, and ③ where to actually buy or have a machine assembled around Yokohama, Japan.
See also our companion article on calling your home-PC AI from smart glasses (Rokid Glasses and others) → Smart Glasses Comparison 2026.
We also built a tool that searches used RTX 3090 PCs across Japanese marketplace listings and judges whether the asking price is below market → Used RTX 3090 PC search & value check (in Japanese, with September 2026 sold-price data).
The short answer
The best first step for most people is a used machine with an RTX 3090 (24GB). As of September 2026 they trade around ¥200,000 in Japan (the Yahoo! Auctions median, up from the ¥130k–180k reported over the summer), run 32B-class models, and the community consensus is still "best value for local LLMs in 2026." With a bigger budget, a new RTX 5090 (32GB); if 16GB is enough for your models, an RTX 5080 or 5070 Ti. Around Yokohama: prebuilt GPU machines at DOSPARA Yokohama and PC Kobo (Pass Kumiai); used bases at Janpara Yokohama and Sofmap Yokohama Vivre; assembly services at Kimura Computer (Aobadai).
How we reviewed this
Criteria: ① usefulness for local LLMs (largest runnable model = VRAM) → ② price/performance → ③ failure resistance (warranty, in-store inspection) → ④ realistic purchase routes in Japan.
Evidence: NVIDIA's official specifications, Japanese technical articles (Zenn/Qiita build write-ups from 2025–2026), market reporting on used prices (as of August 2026), and official shop pages or press coverage (checked September 2026).
Hands-on testing: none claimed. We did not benchmark these GPUs ourselves; speed and price figures are synthesized from public sources.
Updates: First published 2026-09-23 (English). GPU prices move fast — check current listings before buying.
What limits a local LLM, in order: ① VRAM capacity → ② memory bandwidth → ③ compute. If the capacity is too small the model will not fit at all; if it barely fits but bandwidth is narrow, generation crawls. This ordering is a standard observation in hardware-selection write-ups (e.g. Qiita's "choose by capacity, bandwidth, MoE and TTFT").
Rough guide (assuming 4-bit quantization):
| VRAM | Typical models | Experience |
|---|---|---|
| 8GB | 7–8B class | Entry level. Fast but not bright |
| 16GB | 13–14B, up to ~20B with effort | Practical minimum. The 5080/5070 Ti wall |
| 24GB | 20–32B class | Where models start feeling smart. Used-3090 territory |
| 32GB | 32B comfortable; 70B at Q4 barely | Even a 5090 cannot run 70B at full precision (Qiita 2026-04) |
Because smarter models consume more memory, the first decision is the model size you want to run, not your budget. Conversely, 32GB is overkill if 8B covers your use case.
Best value · our pick
Used ~¥200,000 (Sept 2026 Yahoo! Auctions median, ¥185k–210k)
Still widely called the best-value card for local LLMs in 2026 (xda-developers 2026-03, localaimaster 2026-06). 24GB fits 32B-class models; performance is roughly RTX 5070-class. Two of them (48GB) can target 70B-class models.
Downsides: Used market only. You must judge mining history and epoxy aging (see jisaku.com inspection guides). Warranty is shop-limited. Power draw and heat are high; an 850W+ PSU is recommended.
Maximum speed
New and expensive (still scarce in 2026)
32GB VRAM and top-tier bandwidth make it fastest within its model range. There are documented NVFP4 + TensorRT-LLM speedups (Zenn). For people who can solve budget, a ~1000W PSU and heat exhaust all at once.
Downsides: The price is simply very high (Zenn build articles). Even 32GB cannot hold a 70B at full precision; it only fits at Q4, and inference runs at less than a fifth of 7B speed (Qiita 2026-04). Some SKUs are ~55cm long, complicating case choice.
If 16GB is enough
New, mid-high range
Fine up to 13–14B-class models. The CUDA-core difference between the 5070 Ti and 5080 is about 15% against a price gap of roughly ¥100,000, so within 16GB it is largely preference (Zenn).
Downsides: The 16GB wall is hard. If you later want 32B-class models, there is no path except replacement — weaker future-proofing than a used 3090.
The cheapest route is a used prebuilt tower (12th/13th-gen Core i7 or Ryzen 7, etc.) with the GPU and PSU swapped. Four things to check at a used-PC shop:
Inside used GPUs
Cards previously used for mining may have worn fan bearings and heat-degraded components. For shop stock, choose items with an operational warranty, preferably heavy triple-fan models. Popular models with replacement fans still available (EVGA/ASUS/MSI 3090s, etc.) remain repairable later.
All around Yokohama Station and within the city. Presence and stock confirmed from official pages and press coverage as of September 2026. Call or check the web for hours and stock before visiting. (This section is Japan-local by nature; prices in yen, links in Japanese.)
| Shop | Location | Role | Notes |
|---|---|---|---|
| Janpara Yokohama | 2 min from Yokohama Station west (Sotetsu) exit (Nishi-ku Minamisaiwai 1-5-39) | Used PCs, GPUs and parts | Major used-gear chain with shop warranty. A rare place to inspect used GPUs in person |
| Bic Camera Outlet × Sofmap Yokohama Vivre | Yokohama Vivre 7F (station west exit) | Used / outlet PCs | Solid warranty coverage. One of the nearest options for a used-PC base |
| DOSPARA Yokohama Ekimae | Near Yokohama Station (Nishi-ku Minamisaiwai 1-5-30) | Prebuilts (GALLERIA etc.), parts, used | Renewed with stronger parts/used sections (reported by Window's Forest). Demo gaming PCs on display. The place to just buy a GPU-equipped machine |
| PC Kobo (Unitcom BTO + support) | Stores near Yokohama + web BTO | BTO and upgrade installation | 1-year free warranty on BTO (parts + labor). Paid GPU installation/upgrade work accepted through official support |
| Kimura Computer Yokohama (Aobadai) | Aobadai | Assembly service, case swaps | A specialist that assembles customer-supplied parts. For "I won't build it myself, but I choose the parts" |
| PC Val (PC Repair 24) Yokohama Kannai | Kannai | Used PCs, repair | One station away. An option for bargain hunting |
Three buying patterns. A: order a GPU-equipped prebuilt from DOSPARA / PC Kobo (safe, more expensive). B: buy a used tower at Janpara / Sofmap and install the GPU and PSU yourself (cheap, more work). C: gather parts and have Kimura Computer assemble them (the middle ground). For a first build, A or C. B suits people who understand warranties and PSUs.