{"key":"local-builds","locale":"ru","title":"Сборки ПК под локальные нейросети 20B–120B — 2026","url":"https://fedi.software/ru/ai/infographics/local-builds","fields":[{"title":"Сборка","key":"title","format":"text","visible":true,"unit":null},{"title":"Бюджет","key":"tier","format":"badge","visible":true,"unit":null},{"title":"Видеопамять","key":"vram","format":"gb","visible":true,"unit":null},{"title":"ОЗУ","key":"ram","format":"gb","visible":true,"unit":null},{"title":"Цена","key":"price","format":"usd","visible":true,"unit":null},{"title":"Комплектующие","key":"components","format":"list","visible":true,"unit":null},{"title":"Что тянет","key":"models","format":"tok_s","visible":true,"unit":null},{"title":"Примечания","key":"notes","format":"text","visible":true,"unit":null}],"rows":[{"_id":"1x-16gb","title":"Одна видеокарта на 16 ГБ","tier":"Бюджетная","vram":16,"ram":32,"price":null,"components":["Видеокарта: RTX 5060 Ti 16 GB / RTX 4060 Ti 16 GB","Процессор: any 6–8-core desktop CPU","Материнская плата: any ATX/mATX with a PCIe x16 slot","ОЗУ: 32 GB DDR4/DDR5","Блок питания: 550–650 W","PCIe: 1× x16 (x8 electrical is enough)"],"models":[{"id":"gpt-oss-20b","name":"gpt-oss-20b MXFP4 (RTX 5060 Ti 16 GB)","href":"https://fedi.software/ru/ai/compare?m=gpt-oss-20b","quant":"MXFP4","tok_s":111.5,"kind":"measured","source_url":"https://github.com/ggml-org/llama.cpp/discussions/15396","source_date":"2025-08"},{"id":"qwen3-30b-a3b","name":"Qwen3-30B-A3B Q4_K_M, experts partly in RAM","href":"https://fedi.software/ru/ai/compare?m=qwen3-30b-a3b","quant":"Q4_K_M","kind":"estimate"}],"notes":"gpt-oss-20b (≈12 ГБ) целиком помещается в VRAM с запасом под длинный контекст. Qwen3-30B-A3B Q4_K_M (≈18,6 ГБ) не влезает: часть экспертов уходит в RAM через --n-cpu-moe, но работать можно — активных параметров всего ~3B. Замер — на RTX 5060 Ti 16 ГБ (llama-bench tg128)."},{"_id":"1x-3090","title":"Одна RTX 3090 24 ГБ (б/у)","tier":"Бюджетная","vram":24,"ram":64,"price":null,"components":["Видеокарта: RTX 3090 24 GB","Процессор: any 6–8-core desktop CPU","Материнская плата: any ATX with a PCIe x16 slot","ОЗУ: 32–64 GB DDR4/DDR5","Блок питания: 750–850 W","PCIe: 1× x16"],"models":[{"id":"gpt-oss-20b","name":"gpt-oss-20b MXFP4","href":"https://fedi.software/ru/ai/compare?m=gpt-oss-20b","quant":"MXFP4","tok_s":161.8,"kind":"measured","source_url":"https://github.com/ggml-org/llama.cpp/discussions/15396","source_date":"2025-08"},{"id":"qwen3-30b-a3b","name":"Qwen3-30B-A3B Q4_K","href":"https://fedi.software/ru/ai/compare?m=qwen3-30b-a3b","quant":"Q4_K_M","tok_s":153.6,"ctx":"4k","kind":"measured","source_url":"https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/","source_date":"2026-03"},{"id":"gemma-3-27b-it","name":"Gemma 3 27B Q4_K_M","href":"https://fedi.software/ru/ai/compare?m=gemma-3-27b-it","quant":"Q4_K_M","tok_s":[31,42],"kind":"estimate"}],"notes":"Самый дешёвый способ получить 24 ГБ быстрой VRAM. Небольшие MoE летают; плотные модели ~27–32B помещаются в Q4 с умеренным контекстом. Цифра для Gemma 3 27B — наша оценка: 936 ГБ/с × 55–75 % эффективности ÷ 16,5 ГБ весов."},{"_id":"hybrid-gpu-ram","title":"GPU 12–24 ГБ + 64–128 ГБ RAM (выгрузка MoE)","tier":"Бюджетная","vram":24,"ram":64,"price":null,"components":["Видеокарта: RTX 3090 24 GB / RTX 4070 12 GB / RTX 3080 Ti 12 GB","Процессор: Core i5-12600K / Core Ultra 7 265K class","Материнская плата: desktop board, both memory channels populated","ОЗУ: 64–128 GB DDR5 (XMP/EXPO on) or DDR4-3600","Блок питания: 750–850 W","PCIe: 1× x16"],"models":[{"id":"openai-gpt-oss-120b","name":"gpt-oss-120b MXFP4 — RTX 3090 + 64 GB DDR5-5200, --n-cpu-moe 24–26","href":"https://fedi.software/ru/ai/compare?m=openai-gpt-oss-120b","quant":"MXFP4","tok_s":[26.2,27.5],"kind":"measured","source_url":"https://github.com/ggml-org/llama.cpp/discussions/15396","source_date":"2026-01-20"},{"id":"openai-gpt-oss-120b","name":"gpt-oss-120b MXFP4 — RTX 4070 12 GB + 64 GB DDR5","href":"https://fedi.software/ru/ai/compare?m=openai-gpt-oss-120b","quant":"MXFP4","tok_s":[25,28],"kind":"measured","source_url":"https://carteakey.dev/blog/local-inference/optimizing-gpt-oss-120b-local-inference/","source_date":"2025-09-21"},{"id":"openai-gpt-oss-120b","name":"gpt-oss-120b MXFP4 — RTX 3080 Ti 12 GB + 128 GB DDR4-3600, --n-cpu-moe 36","href":"https://fedi.software/ru/ai/compare?m=openai-gpt-oss-120b","quant":"MXFP4","tok_s":[18,22],"kind":"measured","source_url":"https://github.crookster.org/running-gpt-oss-120b-on-rtx-3080-ti-12-gb-at-home/","source_date":"2025-08-25"}],"notes":"llama.cpp --n-cpu-moe N оставляет внимание и общие слои на GPU, а экспертов N слоёв переносит в RAM. Хорошо работает только с MoE; важна скорость памяти — сборка с 4070 выдавала 10–11 ток/с, пока не включили XMP."},{"_id":"2x-3090","title":"Две RTX 3090 (48 ГБ)","tier":"Средняя","vram":48,"ram":64,"price":null,"components":["Видеокарта: 2× RTX 3090 24 GB","Процессор: desktop CPU with x8/x8 bifurcation, or HEDT","Материнская плата: two x16-size slots spaced for 3-slot cards (x8/x8)","ОЗУ: 64 GB","Блок питания: 1000–1200 W","PCIe: 2× x8 PCIe 4.0; NVLink optional"],"models":[{"id":"llama-3-70b-instruct","name":"Llama 3 70B Q4_K_M (same size as Llama 3.3 70B)","href":"https://fedi.software/ru/ai/compare?m=llama-3-70b-instruct","quant":"Q4_K_M","tok_s":16.3,"kind":"measured","source_url":"https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference","source_date":"2024-05"}],"notes":"48 ГБ вмещают плотную 70B в Q4_K_M (≈42,5 ГБ) с коротким или средним контекстом. llama.cpp делит слои между картами и считает их по очереди, поэтому скорость — как у одной 3090, читающей всю модель; NVLink и линии x8 здесь почти не важны (важны для tensor parallel в vLLM)."},{"_id":"epyc-ddr4","title":"Б/у сервер EPYC, 8 каналов DDR4 (только CPU)","tier":"Средняя","vram":0,"ram":512,"price":2000,"components":["Процессор: AMD EPYC 7002/7003 (Rome/Milan), e.g. EPYC 7702","Материнская плата: ASRock Rack ROMED8-2T / Supermicro H12SSL-i / Gigabyte MZ32-AR0","ОЗУ: 256–512 GB DDR4-3200 RDIMM, all 8 channels populated","Блок питания: 750–1000 W","PCIe: 5–7× x16 PCIe 4.0 — room to add GPUs later"],"models":[{"id":"deepseek-r1","name":"DeepSeek R1 671B Q4 — EPYC 7702 + 512 GB DDR4-2400, MZ32-AR0","href":"https://fedi.software/ru/ai/compare?m=deepseek-r1","quant":"Q4","tok_s":[3.5,4.25],"kind":"measured","source_url":"https://digitalspaceport.com/how-to-run-deepseek-r1-671b-fully-locally-on-2000-epyc-rig/","source_date":"2025-01-29"},{"id":"openai-gpt-oss-120b","name":"gpt-oss-120b MXFP4","href":"https://fedi.software/ru/ai/compare?m=openai-gpt-oss-120b","quant":"MXFP4","tok_s":[12,16],"kind":"estimate"}],"notes":"Огромная память за небольшие деньги: 8 каналов DDR4-3200 дают 204,8 ГБ/с на сокет. Практичны только MoE — скорость идёт по активным параметрам. Заполняйте все каналы; второй сокет добавляет проблем с NUMA и мало скорости в llama.cpp (--numa distribute). Цифра для gpt-oss-120b — оценка, пересчитанная по пропускной способности памяти с замера на EPYC с DDR5. Сборка из источника стоила около $2000 (цены января 2025)."},{"_id":"strix-halo","title":"Мини-ПК на AMD Strix Halo, 128 ГБ","tier":"Средняя","vram":96,"ram":128,"price":null,"components":["Видеокарта: Radeon 8060S (integrated)","Процессор: AMD Ryzen AI Max+ 395","Материнская плата: Framework Desktop / other Strix Halo mini-PCs","ОЗУ: 128 GB LPDDR5X-8000 unified (≈96 GB+ usable as VRAM)","Блок питания: built-in"],"models":[{"id":"openai-gpt-oss-120b","name":"gpt-oss-120b MXFP4 (ROCm)","href":"https://fedi.software/ru/ai/compare?m=openai-gpt-oss-120b","quant":"MXFP4","tok_s":50.2,"kind":"measured","source_url":"https://quozul.dev/posts/2025-10-27-gpt-oss-120b-on-ryzen-ai-max-395/","source_date":"2025-10-27"}],"notes":"Тихо, ~120 Вт, вся модель в общей памяти. Пропускная способность (256 ГБ/с) — около четверти от 3090, поэтому плотные 70B медленные; его сильная сторона — большие MoE с малым числом активных параметров."},{"_id":"3x-3090","title":"Три RTX 3090 (72 ГБ)","tier":"Топовая","vram":72,"ram":64,"price":null,"components":["Видеокарта: 3× RTX 3090 24 GB","Процессор: AMD EPYC 7002/7003 or Threadripper (enough PCIe lanes)","Материнская плата: ASRock Rack ROMED8-2T / Supermicro H12SSL-i, risers for 3-slot cards","ОЗУ: 64–128 GB","Блок питания: 1200–1600 W (or two PSUs)","PCIe: 3× x8–x16 PCIe 4.0"],"models":[{"id":"openai-gpt-oss-120b","name":"gpt-oss-120b MXFP4, all in VRAM","href":"https://fedi.software/ru/ai/compare?m=openai-gpt-oss-120b","quant":"MXFP4","tok_s":[41.1,73.3],"ctx":"12.8k→93.7k","kind":"measured","source_url":"https://hardware-corner.net/3x-rtx-3090-gpt-oss-120b-test/","source_date":"2025-10-19"}],"notes":"gpt-oss-120b (≈65 ГБ) целиком помещается в 72 ГБ VRAM с длинным контекстом: 73 ток/с при контексте 12,8k, 41 ток/с при 93,7k. Замер — на арендованной машине, поэтому платформа — наша рекомендация: серверная плата даёт каждой карте достаточно линий. Трём картам по 350 Вт нужен мощный БП; многие ограничивают им лимит мощности."},{"_id":"dgx-spark","title":"NVIDIA DGX Spark, 128 ГБ","tier":"Топовая","vram":128,"ram":128,"price":3999,"components":["Видеокарта: NVIDIA GB10 (integrated Blackwell)","Процессор: Arm CPU of the GB10","Материнская плата: NVIDIA DGX Spark","ОЗУ: 128 GB LPDDR5X unified","Блок питания: built-in"],"models":[{"id":"openai-gpt-oss-120b","name":"gpt-oss-120b MXFP4","href":"https://fedi.software/ru/ai/compare?m=openai-gpt-oss-120b","quant":"MXFP4","tok_s":60.6,"kind":"measured","source_url":"https://github.com/ggml-org/llama.cpp/discussions/16578","source_date":"2025-11-02"}],"notes":"Коробка с CUDA и 128 ГБ общей памяти на 273 ГБ/с: тот же класс скорости, что у Strix Halo, но со стеком NVIDIA. Ранние сборки llama.cpp давали ~35 ток/с на gpt-oss-120b; цифра — после обновления ноября 2025 (llama-bench tg128)."},{"_id":"mac-ultra","title":"Mac Studio на M2/M3 Ultra","tier":"Топовая","vram":192,"ram":192,"price":null,"components":["Видеокарта: Apple M2 Ultra 76-core / M3 Ultra 80-core GPU","Процессор: Apple M2 Ultra / M3 Ultra","Материнская плата: Mac Studio","ОЗУ: 192–512 GB unified (800–819 GB/s)","Блок питания: built-in"],"models":[{"id":"openai-gpt-oss-120b","name":"gpt-oss-120b MXFP4 — M2 Ultra 192 GB","href":"https://fedi.software/ru/ai/compare?m=openai-gpt-oss-120b","quant":"MXFP4","tok_s":79.7,"kind":"measured","source_url":"https://github.com/ggml-org/llama.cpp/discussions/15396","source_date":"2025-08"},{"id":"gpt-oss-20b","name":"gpt-oss-20b MXFP4 — M3 Ultra 512 GB","href":"https://fedi.software/ru/ai/compare?m=gpt-oss-20b","quant":"MXFP4","tok_s":115.5,"kind":"measured","source_url":"https://github.com/ggml-org/llama.cpp/discussions/15396","source_date":"2025-08"}],"notes":"Самый простой способ запустить самые большие модели дома: у M3 Ultra до 512 ГБ общей памяти — хватает на MoE класса DeepSeek в 4 битах. Обработка промпта заметно медленнее, чем на GPU NVIDIA. Для очень больших моделей поднимите лимит памяти GPU через sysctl iogpu.wired_limit_mb."}],"meta":{"sources":[{"name":"github.com","url":"https://github.com/ggml-org/llama.cpp/discussions/15396","license":null,"license_url":null,"modified":false},{"name":"www.hardware-corner.net","url":"https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-3090/","license":null,"license_url":null,"modified":false},{"name":"carteakey.dev","url":"https://carteakey.dev/blog/local-inference/optimizing-gpt-oss-120b-local-inference/","license":null,"license_url":null,"modified":false},{"name":"github.crookster.org","url":"https://github.crookster.org/running-gpt-oss-120b-on-rtx-3080-ti-12-gb-at-home/","license":null,"license_url":null,"modified":false},{"name":"hardware-corner.net","url":"https://hardware-corner.net/3x-rtx-3090-gpt-oss-120b-test/","license":null,"license_url":null,"modified":false},{"name":"digitalspaceport.com","url":"https://digitalspaceport.com/how-to-run-deepseek-r1-671b-fully-locally-on-2000-epyc-rig/","license":null,"license_url":null,"modified":false},{"name":"quozul.dev","url":"https://quozul.dev/posts/2025-10-27-gpt-oss-120b-on-ryzen-ai-max-395/","license":null,"license_url":null,"modified":false}],"updated":"2026-10-02","notes":["Замеры скорости — со ссылкой на источник; ≈ — оценка по пропускной способности памяти. Цены примерные."],"license":{"name":"fedi.software","license":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/","scope":"fedi.software data: hardware, builds, glossary, Fit calculations, curation and the compilation; third-party data keeps the licence of its source (sources[])"},"attribution":["Data by fedi.software, CC BY 4.0 — a link to the source page is required."]}}