← Studio Toriumi

Local LLM · Qwen4-27B · Used GPUs in Japan · 2026-09

A machine for uncensored 27B models,
bought used, for as little as possible.

This is a tool for buying a second-hand GPU machine to run local LLMs — uncensored models included — at home. It searches Mercari, Yahoo! Auctions, Yahoo! Flea Market and Rakuma in one go, then lets you paste any listing and tells you whether it's cheap or expensive for what's inside. It covers the V100 32GB, modified RTX 3080 20GB, RTX 3090 / 3090 Ti, 4090, 5090 and the EVO-X2.

The marketplaces are Japanese and ship domestically; outside Japan you'll need a proxy-buying service. Builds up to about ¥500k are compared in the second half of The ¥6M local-LLM rig.

Prices sampled 2026-09-24 from Yahoo! Auctions sold listings / the build advice draws on tips from someone who runs local LLMs / the checker runs entirely in your browser — nothing you type is sent anywhere

The short answer (September 2026, waiting for Qwen4-27B)

Alibaba announced Qwen4-27B on 22 September; the weights and specs aren't out yet. Uncensored builds are expected to follow soon after release. Uncensoring costs a little intelligence, so you'll want to run it at a higher quantization (Q5–Q6), which makes 32GB or more the comfortable line. From there, pick by what you'll use it for:

01Search four marketplaces at once

Pick a keyword and a price range to get links to each site's results, cheapest first and only items still for sale. "Sold prices" shows what things actually went for. Keep the keywords in Japanese — that's how sellers write them.

These are plain links to each site's own search page. This site doesn't fetch or store anything from them.

02Paste a listing, get a verdict

Copy the title, description and price from a listing, paste it here and hit "Read listing". Fix any field it gets wrong by hand. Japanese listings work as-is, and it handles both bare cards and complete PCs.

Assumptions (used prices — editable)

Estimated value = the GPU(s). For a complete PC, add the platform (CPU and board), RAM, SSD, PSU and case. The EVO-X2 is priced as a whole machine. A value-to-price ratio of 1.20 or more is a great deal, 1.00–1.20 is around market, and below that is overpriced. GPU and EVO-X2 figures are September 2026 Yahoo! Auctions medians (the 3080 20GB is a rough guide — there are very few sales); the rest are typical private-sale prices for used parts. Change them when the market moves.

03Which one, for what

How fast an LLM writes is set mostly by memory bandwidth, not raw compute. Below is the simple ceiling for a 27B model at Q4 (about 17 GB — bandwidth ÷ 17 GB) next to September 2026 second-hand prices.

HardwareVRAMBandwidth27B Q4 ceilingUsed pricePer GBBest for
Tesla V100 32GB32GB900GB/s~53 tok/s¥108,000~¥3,400Chat. The cheapest 32GB
RTX 3080 20GB (modified)20GB760GB/s~45 tok/s¥69,000 (rough)~¥3,500Two cards for 40GB
RTX 309024GB936GB/s~55 tok/s¥200,000~¥8,300All-rounder: LLMs and video
RTX 409024GB1,008GB/s~59 tok/s¥428,000~¥17,800Overpriced for LLMs alone
RTX 509032GB1,792GB/s~105 tok/s¥910,000~¥28,400Fastest, if money is no object
EVO-X2 64GB64GB (shared)256GB/s~15 tok/s¥249,000 (whole box)~¥3,900No CUDA, big MoE models, silence

Not sure yet? You can rent a comparable GPU by the hour and check whether the model you want is good enough for your use before buying — and avoid a machine you end up barely using.

04Sizing for Qwen4-27B

Qwen4-27B's specs aren't public yet. This table assumes it's a dense 27B model like the recent Qwen 27B releases and works out the weight size at each quantization from bits per parameter. Conversation history (the KV cache) needs a few GB on top.

QuantSize24GB
3090
32GB
V100, 5090
40–48GB
two cards
64GB
EVO-X2
Q4_K_M~16.5GBComfortableComfortableComfortableFits (slow)
Q5_K_M~19GBFits (shorter context)ComfortableComfortableFits (slow)
Q6_K~22GBJust barelyFitsComfortableFits (slow)
Q8_0~29GBDoesn't fitJust barelyComfortableFits (slow)

Uncensored models (made by abliteration or similar pruning) lose a little intelligence compared with the original. You don't want to lose more to heavy quantization, so 32GB for Q5–Q6 is the comfortable line, and 40–48GB across two cards if you want Q8 with long context. A single 24GB 3090 still runs Q4–Q5 well. If Qwen4-27B turns out to be a mixture-of-experts model rather than dense, its total size will grow, and the EVO-X2's 64–128GB comes into its own.

05Market snapshot (2026-09-24)

Yahoo! Auctions sold listings, excluding junk, untested units, multi-card bundles, and absurdly cheap box-only sales. These are prices things actually sold for, not asking prices.

WhatSales · periodMedianRange
Tesla V100 32GB (mostly with fan)45
2026-09-04 – 09-24
¥108,000¥108,000 – 125,000
HP Z440 + V100 32GB bundle2
2026-09
—¥128,000 and ¥178,000
RTX 3080 20GB (modified)2
2026-04 – 06
¥69,000A junk unit went for ¥36,000
RTX 3090 card42
2026-08-29 – 09-23
¥200,000¥185,000 – 210,000
Desktop with 3090 / 3090 Ti16
2026-07-12 – 09-21
¥239,000¥202,000 – 261,000
EVO-X2 64GB8
2026-03-29 – 09-02
¥249,000¥233,000 – 268,000
EVO-X2 128GB3 in Sept / 13 in Apr–Aug¥500,000 / ¥400,000Up about 25% in September
RTX 4090 card32
2026-08-01 – 09-23
¥428,000¥410,000 – 440,000
Desktop with 409015
2026-07-17 – 09-22
¥520,000¥493,000 – 574,000
RTX 5090 card (September)18
2026-09-01 – 09-23
¥910,000¥745,000 in late August — up 20% in a month

Since early September, one seller has been steadily listing V100s "with fan and cable, tested" in the low ¥100k range. With bare 3090s this expensive, a complete 3090 PC gets you the rest of the machine for about ¥40k more. Mercari only renders results in a browser, so it isn't in the numbers — use the links above to check it directly.

06Checklist for buying from private sellers

07What to pick up once it arrives

None of the links on this English page are affiliate links. The Japanese page links to store searches on Amazon and Rakuten with affiliate tags; the marketplace links are never affiliate links, and verdicts don't depend on referral fees.

08Method and sources

  1. Yahoo! Auctions sold-price search (auctions.yahoo.co.jp/closedsearch) — "V100 32GB" latest 50; "3080 20GB"; "RTX3090 デスクトップ" 43 sales over 180 days; "RTX3090 24GB" latest 50; "EVO-X2" 32; "RTX4090 デスクトップ" 41; "RTX4090 24GB" latest 50; "RTX5090 32GB" latest 50 at ¥300k+. Retrieved 2026-09-23 to 24. Junk, untested, bundles, complete PCs (in card rows) and box-only listings removed before taking medians.
  2. Qwen4 series (Max, Plus, Flash, 27B) announced at Alibaba's Apsara Conference on 2026-09-22 — no weights, release date or specs yet (Pandaily, Yotta Labs).
  3. Official memory bandwidth — Tesla V100 900 GB/s (HBM2), RTX 3080 760 GB/s, 3090 936 GB/s, 3090 Ti and 4090 1,008 GB/s, 5090 1,792 GB/s, Ryzen AI Max+ 395 (EVO-X2) 256 GB/s.
  4. Build advice (blower-fan V100s, worn 3090s, modified 3080 20GB pairs, the EVO-X2, RAM for video generation) — tips from someone who runs local LLMs (2026-09). The V100 speed figure is from the same person.
  5. Weight sizes by quantization — estimated from bits per parameter (Q4_K_M ~4.85, Q5_K_M ~5.7, Q6_K ~6.6, Q8_0 ~8.5).