# Best Local AI Rig

> Independent price and spec tracker for hardware that runs open-weight LLMs locally (NVIDIA DGX Spark and GB10 boxes, AMD Strix Halo mini PCs, Apple Mac Studio/Mac mini, NVIDIA RTX GPUs). For each machine: today's retailer price and stock, memory and bandwidth, and which of 22 open models fit at 4-bit/8-bit with estimated tokens/sec. Run by the Best Local AI Rig Team; contact hello@bestlocalairig.com. Updated 2026-10-24.

Fit and speed figures are estimates (formula on /methodology/, calibrated against public llama.cpp/MLX benchmarks), not our own lab tests. Prices change daily; each price shows its check date.

## Start here

- [Best hardware for local LLMs (2026): prices & what runs](https://bestlocalairig.com/): Compare DGX Spark, Strix Halo mini PCs, Mac Studio and RTX GPUs for running LLMs locally: live prices, stock, memory bandwidth and which open models fit.
- [Methodology: how we estimate LLM fit & tokens/sec](https://bestlocalairig.com/methodology/): How Best Local AI Rig estimates which LLMs fit each machine and how fast they run, the public benchmarks we calibrate against, our price sources and update…
- [About Best Local AI Rig: who we are & how we work](https://bestlocalairig.com/about/): Best Local AI Rig is an independent price and spec tracker for local-AI hardware, run by the Best Local AI Rig Team.
- [Local LLM hardware guide (2026): VRAM, models & budgets](https://bestlocalairig.com/guides/): What fits in 8GB to 128GB, the best hardware for each open model, the best rigs by budget, and Mac Studio vs Strix Halo vs DGX Spark.
- [LLM VRAM requirements by model: 4-bit & 8-bit (2026)](https://bestlocalairig.com/models/): Memory needed at 4-bit and 8-bit for 22 open models (Llama, Qwen, DeepSeek R1 distills, Gemma, GPT-OSS, Mistral) and the cheapest rig that runs each.
- [Local AI hardware comparisons](https://bestlocalairig.com/compare/): DGX Spark vs Strix Halo vs Mac Studio vs RTX: side-by-side price, memory, bandwidth and model fit.
- [LLM tokens per second: our estimates vs 52 real benchmarks](https://bestlocalairig.com/guides/speed-benchmarks/): Our tokens/sec estimates vs 52 published llama.cpp and MLX runs on RTX, DGX Spark, Strix Halo and Mac. Median error went from 26% to 5%.

## Buying guides

- [What LLMs can I run with 8GB VRAM? (2026)](https://bestlocalairig.com/guides/what-can-i-run-on-8gb-vram/): 1 open models fit in 8GB VRAM, up to Qwen3.5 9B. Quant, memory needed and estimated tokens/sec, plus the cheapest rigs with more memory.
- [What LLMs can I run with 12GB VRAM? (2026)](https://bestlocalairig.com/guides/what-can-i-run-on-12gb-vram/): 2 open models fit in 12GB VRAM, up to Gemma 4 12B. Quant, memory needed and estimated tokens/sec, plus the cheapest rigs with more memory.
- [What LLMs can I run with 16GB VRAM? (2026)](https://bestlocalairig.com/guides/what-can-i-run-on-16gb-vram/): 3 open models fit in 16GB VRAM, up to GPT-OSS 20B. Quant, memory needed and estimated tokens/sec, plus the cheapest rigs with more memory.
- [What LLMs can I run with 24GB VRAM? (2026)](https://bestlocalairig.com/guides/what-can-i-run-on-24gb-vram/): 7 open models fit in 24GB VRAM, up to Gemma 4 31B. Quant, memory needed and estimated tokens/sec, plus current prices for 1 24GB rig.
- [What LLMs can I run with 32GB VRAM? (2026)](https://bestlocalairig.com/guides/what-can-i-run-on-32gb-vram/): 8 open models fit in 32GB VRAM, up to Qwen3.6 35B-A3B. Quant, memory needed and estimated tokens/sec, plus current prices for 1 32GB rig.
- [What LLMs can I run with 64GB unified memory? (2026)](https://bestlocalairig.com/guides/what-can-i-run-on-64gb-unified-memory/): 9 open models fit in 64GB unified memory, up to Llama 3.3 70B. Quant, memory needed and estimated tokens/sec, plus current prices for 1 64GB rig.
- [What LLMs can I run with 96GB unified memory? (2026)](https://bestlocalairig.com/guides/what-can-i-run-on-96gb-unified-memory/): 12 open models fit in 96GB unified memory, up to GPT-OSS 120B. Quant, memory needed and estimated tokens/sec, plus current prices for 1 96GB rig.
- [What LLMs can I run with 128GB unified memory? (2026)](https://bestlocalairig.com/guides/what-can-i-run-on-128gb-unified-memory/): 14 open models fit in 128GB unified memory, up to Qwen3.5 122B-A10B. Quant, memory needed and estimated tokens/sec, plus current prices for 9 128GB rigs.
- [Best local AI rig under $2,000 (2026): prices & what runs](https://bestlocalairig.com/guides/best-local-ai-rig-under-2000/): 1 local-AI machines under $2,000 right now, ranked by usable memory and bandwidth. Top pick: Mac mini M5 Pro at $1,699. With the open models each one runs.
- [Best local AI rig under $2,500 (2026): prices & what runs](https://bestlocalairig.com/guides/best-local-ai-rig-under-2500/): 3 local-AI machines under $2,500 right now, ranked by usable memory and bandwidth. Top pick: GMKtec EVO-X2 at $2,200. With the open models each one runs.
- [Best local AI rig under $3,500 (2026): prices & what runs](https://bestlocalairig.com/guides/best-local-ai-rig-under-3500/): 5 local-AI machines under $3,500 right now, ranked by usable memory and bandwidth. Top pick: GMKtec EVO-X2 at $3,500. With the open models each one runs.
- [Best local AI rig under $5,000 (2026): prices & what runs](https://bestlocalairig.com/guides/best-local-ai-rig-under-5000/): 9 local-AI machines under $5,000 right now, ranked by usable memory and bandwidth. Top pick: GMKtec EVO-X2 at $3,500. With the open models each one runs.
- [Mac Studio vs Strix Halo for local LLMs (2026): speed & price](https://bestlocalairig.com/guides/mac-studio-vs-strix-halo/): Mac Studio vs Strix Halo for running LLMs at home: current price ranges, memory, bandwidth and which open models each runs, with estimated tokens/sec.
- [DGX Spark vs Strix Halo for local LLMs (2026): speed & price](https://bestlocalairig.com/guides/dgx-spark-vs-strix-halo/): DGX Spark vs Strix Halo for running LLMs at home: current price ranges, memory, bandwidth and which open models each runs, with estimated tokens/sec.
- [DGX Spark vs Mac Studio for local LLMs (2026): speed & price](https://bestlocalairig.com/guides/dgx-spark-vs-mac-studio/): DGX Spark vs Mac Studio for running LLMs at home: current price ranges, memory, bandwidth and which open models each runs, with estimated tokens/sec.
- [Can DGX Spark run Llama 70B? Speed & memory (2026)](https://bestlocalairig.com/guides/can-dgx-spark-run-llama-70b/): Yes: Llama 3.3 70B needs ~46GB at 4-bit and DGX Spark has ~109GB usable, but expect only ~5 tok/s. Measured runs, Q8 fit and faster MoE alternatives.
- [Best mini PC for local LLMs under $2,000 (2026)](https://bestlocalairig.com/guides/best-mini-pc-for-local-llms-under-2000/): The mini PCs for running LLMs at home under $2,000, ranked by usable memory. Top pick: Mac mini M5 Pro at $1,699.
- [Best GPU for local LLMs (2026): VRAM, speed & prices](https://bestlocalairig.com/guides/best-gpu-for-local-llms/): RTX 4090 vs RTX 5090 vs RTX PRO 6000 for running LLMs at home: current prices, VRAM, bandwidth and which open models each card fits, with estimated and…
- [How much VRAM do you need for local LLMs? (2026)](https://bestlocalairig.com/guides/how-much-memory-to-run-llms-locally/): Memory needed for Llama 3.3 70B (~46GB at 4-bit), Qwen, Gemma and GPT-OSS models at 4-bit and 8-bit, what 8GB to 128GB runs, and the cheapest rig that fits…
- [Cheapest way to run a 70B LLM locally (2026)](https://bestlocalairig.com/guides/cheapest-way-to-run-70b-locally/): Every rig we track that fits Llama 3.3 70B, cheapest first: from $2,200 (GMKtec EVO-X2), with estimated and measured tokens/sec and price per tok/s.
- [Is Strix Halo good for local LLMs? Benchmarks & prices (2026)](https://bestlocalairig.com/guides/is-strix-halo-good-for-local-llms/): What a 128GB AMD Strix Halo (Ryzen AI Max+ 395) mini PC runs: GPT-OSS 120B ~55 tok/s, 70B ~5 tok/s.
- [DGX Spark vs RTX 5090 for local LLMs (2026)](https://bestlocalairig.com/guides/dgx-spark-vs-rtx-5090/): DGX Spark (128GB, 273GB/s) vs RTX 5090 (32GB, 1,792GB/s): which open models each fits, estimated and measured tokens/sec, and current prices.

## Best hardware per model

- [Best hardware to run Qwen3.5 9B locally (2026)](https://bestlocalairig.com/run/qwen3-5-9b/): Qwen3.5 9B needs ~7GB at 4-bit. Cheapest rig that fits: Mac mini M5 Pro at $1,699. Best value and fastest picks, quant advice and est. tokens/sec.
- [Best hardware to run Gemma 4 12B locally (2026)](https://bestlocalairig.com/run/gemma-4-12b/): Gemma 4 12B needs ~9GB at 4-bit. Cheapest rig that fits: Mac mini M5 Pro at $1,699. Best value and fastest picks, quant advice and est. tokens/sec.
- [Best hardware to run GPT-OSS 20B locally (2026)](https://bestlocalairig.com/run/gpt-oss-20b/): GPT-OSS 20B needs ~13GB at 4-bit. Cheapest rig that fits: Mac mini M5 Pro at $1,699. Best value and fastest picks, quant advice and est. tokens/sec.
- [Best hardware to run Gemma 4 26B-A4B locally (2026)](https://bestlocalairig.com/run/gemma-4-26b-a4b/): Gemma 4 26B-A4B needs ~18GB at 4-bit. Cheapest rig that fits: GMKtec EVO-X2 at $2,200. Best value and fastest picks, quant advice and est. tokens/sec.
- [Best hardware to run Qwen3.8 27B locally (2026)](https://bestlocalairig.com/run/qwen3-8-27b/): Qwen3.8 27B needs ~19GB at 4-bit. Cheapest rig that fits: GMKtec EVO-X2 at $2,200. Best value and fastest picks, quant advice and est. tokens/sec.
- [Best hardware to run GLM-4.7-Flash 30B-A3B locally (2026)](https://bestlocalairig.com/run/glm-4-7-flash/): GLM-4.7-Flash 30B-A3B needs ~21GB at 4-bit. Cheapest rig that fits: GMKtec EVO-X2 at $2,200. Best value and fastest picks, quant advice and est. tokens/sec.
- [Best hardware to run Gemma 4 31B locally (2026)](https://bestlocalairig.com/run/gemma-4-31b/): Gemma 4 31B needs ~21GB at 4-bit. Cheapest rig that fits: GMKtec EVO-X2 at $2,200. Best value and fastest picks, quant advice and est. tokens/sec.
- [Best hardware to run Qwen3.6 35B-A3B locally (2026)](https://bestlocalairig.com/run/qwen3-6-35b-a3b/): Qwen3.6 35B-A3B needs ~24GB at 4-bit. Cheapest rig that fits: GMKtec EVO-X2 at $2,200. Best value and fastest picks, quant advice and est. tokens/sec.
- [Best hardware to run Llama 3.3 70B locally (2026)](https://bestlocalairig.com/run/llama-3-3-70b/): Llama 3.3 70B needs ~46GB at 4-bit. Cheapest rig that fits: GMKtec EVO-X2 at $2,200. Best value and fastest picks, quant advice and est. tokens/sec.
- [Best hardware to run Qwen3-Next 80B-A3B locally (2026)](https://bestlocalairig.com/run/qwen3-next-80b-a3b/): Qwen3-Next 80B-A3B needs ~52GB at 4-bit. Cheapest rig that fits: GMKtec EVO-X2 at $3,500. Best value and fastest picks, quant advice and est. tokens/sec.
- [Best hardware to run Llama 4 Scout 109B-A17B locally (2026)](https://bestlocalairig.com/run/llama-4-scout/): Llama 4 Scout 109B-A17B needs ~71GB at 4-bit. Cheapest rig that fits: GMKtec EVO-X2 at $3,500. Best value and fastest picks, quant advice and est. tokens/sec.
- [Best hardware to run GPT-OSS 120B locally (2026)](https://bestlocalairig.com/run/gpt-oss-120b/): GPT-OSS 120B needs ~67GB at 4-bit. Cheapest rig that fits: GMKtec EVO-X2 at $3,500. Best value and fastest picks, quant advice and est. tokens/sec.
- [Best hardware to run Nemotron 3 Super 120B-A12B locally (2026)](https://bestlocalairig.com/run/nemotron-3-super-120b/): Nemotron 3 Super 120B-A12B needs ~78GB at 4-bit. Cheapest rig that fits: GMKtec EVO-X2 at $3,500. Best value and fastest picks, quant advice and est.
- [Best hardware to run Qwen3.5 122B-A10B locally (2026)](https://bestlocalairig.com/run/qwen3-5-122b-a10b/): Qwen3.5 122B-A10B needs ~79GB at 4-bit. Cheapest rig that fits: GMKtec EVO-X2 at $3,500. Best value and fastest picks, quant advice and est. tokens/sec.

## Memory requirements per model

- [Qwen3.5 9B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/qwen3-5-9b/): Qwen3.5 9B needs ~7GB at 4-bit, ~12GB at 8-bit. 18 rigs fit it, from Mac mini M5 Pro at $1,699. Est. tokens/sec, quant and today's prices.
- [Gemma 4 12B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/gemma-4-12b/): Gemma 4 12B needs ~9GB at 4-bit, ~15GB at 8-bit. 18 rigs fit it, from Mac mini M5 Pro at $1,699. Est. tokens/sec, quant and today's prices.
- [GPT-OSS 20B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/gpt-oss-20b/): GPT-OSS 20B needs ~13GB at 4-bit, ~25GB at 8-bit. 18 rigs fit it, from Mac mini M5 Pro at $1,699. Est. tokens/sec, quant and today's prices.
- [Gemma 4 26B-A4B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/gemma-4-26b-a4b/): Gemma 4 26B-A4B needs ~18GB at 4-bit, ~30GB at 8-bit. 17 rigs fit it, from GMKtec EVO-X2 at $2,200. Est. tokens/sec, quant and today's prices.
- [Qwen3.8 27B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/qwen3-8-27b/): Qwen3.8 27B needs ~19GB at 4-bit, ~32GB at 8-bit. 17 rigs fit it, from GMKtec EVO-X2 at $2,200. Est. tokens/sec, quant and today's prices.
- [Gemma 3 27B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/gemma-3-27b/): Gemma 3 27B needs ~19GB at 4-bit, ~32GB at 8-bit. 17 rigs fit it, from GMKtec EVO-X2 at $2,200. Est. tokens/sec, quant and today's prices.
- [GLM-4.7-Flash 30B-A3B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/glm-4-7-flash/): GLM-4.7-Flash 30B-A3B needs ~21GB at 4-bit, ~35GB at 8-bit. 17 rigs fit it, from GMKtec EVO-X2 at $2,200. Est. tokens/sec, quant and today's prices.
- [Qwen3 30B-A3B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/qwen3-30b-a3b/): Qwen3 30B-A3B needs ~21GB at 4-bit, ~36GB at 8-bit. 17 rigs fit it, from GMKtec EVO-X2 at $2,200. Est. tokens/sec, quant and today's prices.
- [Gemma 4 31B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/gemma-4-31b/): Gemma 4 31B needs ~21GB at 4-bit, ~36GB at 8-bit. 17 rigs fit it, from GMKtec EVO-X2 at $2,200. Est. tokens/sec, quant and today's prices.
- [Qwen3 32B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/qwen3-32b/): Qwen3 32B needs ~22GB at 4-bit, ~38GB at 8-bit. 16 rigs fit it, from GMKtec EVO-X2 at $2,200. Est. tokens/sec, quant and today's prices.
- [DeepSeek R1 Distill Qwen 32B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/deepseek-r1-distill-qwen-32b/): DeepSeek R1 Distill Qwen 32B needs ~22GB at 4-bit, ~38GB at 8-bit. 16 rigs fit it, from GMKtec EVO-X2 at $2,200. Est. tokens/sec, quant and today's prices.
- [Qwen3.6 35B-A3B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/qwen3-6-35b-a3b/): Qwen3.6 35B-A3B needs ~24GB at 4-bit, ~41GB at 8-bit. 16 rigs fit it, from GMKtec EVO-X2 at $2,200. Est. tokens/sec, quant and today's prices.
- [Llama 3.3 70B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/llama-3-3-70b/): Llama 3.3 70B needs ~46GB at 4-bit, ~80GB at 8-bit. 13 rigs fit it, from GMKtec EVO-X2 at $2,200. Est. tokens/sec, quant and today's prices.
- [DeepSeek R1 Distill Llama 70B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/deepseek-r1-distill-llama-70b/): DeepSeek R1 Distill Llama 70B needs ~46GB at 4-bit, ~80GB at 8-bit. 13 rigs fit it, from GMKtec EVO-X2 at $2,200. Est. tokens/sec, quant and today's prices.
- [Qwen3-Next 80B-A3B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/qwen3-next-80b-a3b/): Qwen3-Next 80B-A3B needs ~52GB at 4-bit, ~91GB at 8-bit. 12 rigs fit it, from GMKtec EVO-X2 at $3,500. Est. tokens/sec, quant and today's prices.
- [Llama 4 Scout 109B-A17B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/llama-4-scout/): Llama 4 Scout 109B-A17B needs ~71GB at 4-bit, ~123GB at 8-bit. 12 rigs fit it, from GMKtec EVO-X2 at $3,500. Est. tokens/sec, quant and today's prices.
- [GPT-OSS 120B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/gpt-oss-120b/): GPT-OSS 120B needs ~67GB at 4-bit, ~132GB at 8-bit. 12 rigs fit it, from GMKtec EVO-X2 at $3,500. Est. tokens/sec, quant and today's prices.
- [Mistral Small 4 119B-A6B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/mistral-small-4-119b/): Mistral Small 4 119B-A6B needs ~78GB at 4-bit, ~135GB at 8-bit. 11 rigs fit it, from GMKtec EVO-X2 at $3,500. Est. tokens/sec, quant and today's prices.
- [Nemotron 3 Super 120B-A12B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/nemotron-3-super-120b/): Nemotron 3 Super 120B-A12B needs ~78GB at 4-bit, ~135GB at 8-bit. 11 rigs fit it, from GMKtec EVO-X2 at $3,500. Est. tokens/sec, quant and today's prices.
- [Qwen3.5 122B-A10B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/qwen3-5-122b-a10b/): Qwen3.5 122B-A10B needs ~79GB at 4-bit, ~138GB at 8-bit. 11 rigs fit it, from GMKtec EVO-X2 at $3,500. Est. tokens/sec, quant and today's prices.
- [Qwen3 235B-A22B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/qwen3-235b-a22b/): Qwen3 235B-A22B needs ~151GB at 4-bit, ~264GB at 8-bit. 1 rigs fit it. Est. tokens/sec, quant and prices.
- [Muse Spark Open 816B-A42B VRAM requirements & hardware (2026)](https://bestlocalairig.com/models/muse-spark-open/): Muse Spark Open 816B-A42B needs ~521GB at 4-bit, ~912GB at 8-bit. Too big for any single desktop we track; see what it would take.

## Machines

- [NVIDIA DGX Spark for local LLMs: price, speed & what it runs](https://bestlocalairig.com/machines/nvidia-dgx-spark/): NVIDIA DGX Spark (Founders Edition, 4TB): 128GB at 273 GB/s. Current price $6,950, stock, and estimated fit + tokens/sec for 16 open models.
- [ASUS Ascent GX10 for local LLMs: price, speed & what it runs](https://bestlocalairig.com/machines/asus-ascent-gx10/): ASUS Ascent GX10 (1TB): 128GB at 273 GB/s. Current price $5,999, stock, and estimated fit + tokens/sec for 16 open models.
- [Gigabyte AI TOP ATOM for local LLMs: price & speed](https://bestlocalairig.com/machines/gigabyte-ai-top-atom/): Gigabyte AI TOP ATOM (4TB): 128GB at 273 GB/s. Current price $8,499, stock, and estimated fit + tokens/sec for 16 open models.
- [Lenovo ThinkStation PGX for local LLMs: price & speed](https://bestlocalairig.com/machines/lenovo-thinkstation-pgx/): Lenovo ThinkStation PGX (1TB): 128GB at 273 GB/s. Current price $5,859, stock, and estimated fit + tokens/sec for 16 open models.
- [HP ZGX Nano G1n for local LLMs: price, speed & what it runs](https://bestlocalairig.com/machines/hp-zgx-nano-g1n/): HP ZGX Nano G1n (4TB): 128GB at 273 GB/s. Current price $7,699, stock, and estimated fit + tokens/sec for 16 open models.
- [GMKtec EVO-X2 128GB for local LLMs: price & speed](https://bestlocalairig.com/machines/gmktec-evo-x2-128gb/): GMKtec EVO-X2 (Ryzen AI Max+ 395, 128GB / 1TB): 128GB at 256 GB/s. Current price $3,500, stock, and estimated fit + tokens/sec for 16 open models.
- [GMKtec EVO-X2 64GB for local LLMs: price & speed](https://bestlocalairig.com/machines/gmktec-evo-x2-64gb/): GMKtec EVO-X2 (Ryzen AI Max+ 395, 64GB / 1TB): 64GB at 256 GB/s. Current price $2,200, stock, and estimated fit + tokens/sec for 16 open models.
- [Framework Desktop DIY for local LLMs: price & speed](https://bestlocalairig.com/machines/framework-desktop-128gb/): Framework Desktop DIY (Ryzen AI Max+ 395, 128GB): 128GB at 256 GB/s. Current price $4,839, stock, and estimated fit + tokens/sec for 16 open models.
- [Minisforum MS-S1 Max for local LLMs: price & speed](https://bestlocalairig.com/machines/minisforum-ms-s1-max/): Minisforum MS-S1 Max (Ryzen AI Max+ 395, 128GB / 2TB): 128GB at 256 GB/s. Current price $3,919, stock, and estimated fit + tokens/sec for 16 open models.
- [Beelink GTR9 Pro for local LLMs: price, speed & what it runs](https://bestlocalairig.com/machines/beelink-gtr9-pro/): Beelink GTR9 Pro (Ryzen AI Max+ 395, 128GB / 2TB): 128GB at 256 GB/s. Current price $4,349, stock, and estimated fit + tokens/sec for 16 open models.
- [Mac Studio M5 Max 36GB for local LLMs: price & speed](https://bestlocalairig.com/machines/mac-studio-m5-max-36gb/): Mac Studio M5 Max (18-core CPU / 32-core GPU, 36GB): 36GB at 460 GB/s. Current price $2,499, stock, and estimated fit + tokens/sec for 16 open models.
- [Mac Studio M5 Max 48GB for local LLMs: price & speed](https://bestlocalairig.com/machines/mac-studio-m5-max-48gb/): Mac Studio M5 Max (18-core CPU / 40-core GPU, 48GB): 48GB at 614 GB/s. Current price $3,099, stock, and estimated fit + tokens/sec for 16 open models.
- [Mac Studio M5 Ultra 96GB for local LLMs: price & speed](https://bestlocalairig.com/machines/mac-studio-m5-ultra-96gb/): Mac Studio M5 Ultra (30-core CPU / 64-core GPU, 96GB): 96GB at 1200 GB/s. Current price $5,499, stock, and estimated fit + tokens/sec for 16 open models.
- [Mac Studio M5 Ultra 256GB for local LLMs: price & speed](https://bestlocalairig.com/machines/mac-studio-m5-ultra-256gb/): Mac Studio M5 Ultra (256GB): 256GB at 1200 GB/s. Current price unknown, stock, and estimated fit + tokens/sec for 16 open models.
- [Mac mini M5 Pro for local LLMs: price, speed & what it runs](https://bestlocalairig.com/machines/mac-mini-m5-pro-24gb/): Mac mini M5 Pro (15-core CPU / 16-core GPU, 24GB): 24GB at 307 GB/s. Current price $1,699, stock, and estimated fit + tokens/sec for 16 open models.
- [RTX 5090 FE for local LLMs: price, speed & what it runs](https://bestlocalairig.com/machines/nvidia-rtx-5090-fe/): NVIDIA GeForce RTX 5090 Founders Edition (32GB card): 32GB at 1792 GB/s. Current price $8,500, stock, and estimated fit + tokens/sec for 16 open models.
- [Gigabyte RTX 4090 for local LLMs: price & speed](https://bestlocalairig.com/machines/rtx-4090-gigabyte-windforce/): Gigabyte GeForce RTX 4090 Windforce V2 (24GB card): 24GB at 1008 GB/s. Current price $4,899, stock, and estimated fit + tokens/sec for 16 open models.
- [RTX PRO 6000 for local LLMs: price, speed & what it runs](https://bestlocalairig.com/machines/nvidia-rtx-pro-6000-blackwell/): PNY NVIDIA RTX PRO 6000 Blackwell (96GB card): 96GB at 1792 GB/s. Current price $17,899, stock, and estimated fit + tokens/sec for 16 open models.

## Comparisons

- [NVIDIA DGX Spark vs GMKtec EVO-X2 128GB for local LLMs](https://bestlocalairig.com/compare/nvidia-dgx-spark-vs-gmktec-evo-x2-128gb/): NVIDIA DGX Spark vs GMKtec EVO-X2: current prices, memory, bandwidth and which open models each runs, with estimated tokens/sec.
- [NVIDIA DGX Spark vs Mac Studio M5 Ultra 96GB for local LLMs](https://bestlocalairig.com/compare/nvidia-dgx-spark-vs-mac-studio-m5-ultra-96gb/): NVIDIA DGX Spark vs Mac Studio M5 Ultra: current prices, memory, bandwidth and which open models each runs, with estimated tokens/sec.
- [ASUS Ascent GX10 vs NVIDIA DGX Spark for local LLMs](https://bestlocalairig.com/compare/asus-ascent-gx10-vs-nvidia-dgx-spark/): ASUS Ascent GX10 vs NVIDIA DGX Spark: current prices, memory, bandwidth and which open models each runs, with estimated tokens/sec.
- [Framework Desktop DIY vs GMKtec EVO-X2 128GB for local LLMs](https://bestlocalairig.com/compare/framework-desktop-128gb-vs-gmktec-evo-x2-128gb/): Framework Desktop DIY vs GMKtec EVO-X2: current prices, memory, bandwidth and which open models each runs, with estimated tokens/sec.
- [RTX 5090 FE vs Mac Studio M5 Max 48GB for local LLMs](https://bestlocalairig.com/compare/nvidia-rtx-5090-fe-vs-mac-studio-m5-max-48gb/): NVIDIA GeForce RTX 5090 Founders Edition vs Mac Studio M5 Max: current prices, memory, bandwidth and which open models each runs, with estimated tokens/sec.
- [Minisforum MS-S1 Max vs Beelink GTR9 Pro for local LLMs](https://bestlocalairig.com/compare/minisforum-ms-s1-max-vs-beelink-gtr9-pro/): Minisforum MS-S1 Max vs Beelink GTR9 Pro: current prices, memory, bandwidth and which open models each runs, with estimated tokens/sec.

## Optional

- [Full facts for LLMs](https://bestlocalairig.com/llms-full.txt): key numbers, every machine and every model in plain text
- [Fit matrix JSON](https://bestlocalairig.com/fit-matrix.json): machine × model fit and est. tokens/sec
- [RSS feed](https://bestlocalairig.com/feed.xml)

## Français

Version française du site (mêmes données, prix en dollars américains). Toutes les pages anglaises ont un équivalent sous https://bestlocalairig.com/fr/.

- [IA locale : quel matériel pour faire tourner un LLM en local](https://bestlocalairig.com/fr/)
- [Méthode : compatibilité des LLM et jetons/s](https://bestlocalairig.com/fr/methodology/)
- [Guides matériel pour l’IA locale (2026) : VRAM, LLM, budgets](https://bestlocalairig.com/fr/guides/)
- [VRAM nécessaire par LLM : 4 et 8 bits (2026)](https://bestlocalairig.com/fr/models/)
- [Comparatifs de matériel pour l’IA locale](https://bestlocalairig.com/fr/compare/)
- [Vitesse des LLM : nos estimations vs 52 vrais bancs d’essai](https://bestlocalairig.com/fr/guides/speed-benchmarks/)
- [Quels LLM avec 8 Go de VRAM? (2026)](https://bestlocalairig.com/fr/guides/what-can-i-run-on-8gb-vram/)
- [Quels LLM avec 12 Go de VRAM? (2026)](https://bestlocalairig.com/fr/guides/what-can-i-run-on-12gb-vram/)
- [Quels LLM avec 16 Go de VRAM? (2026)](https://bestlocalairig.com/fr/guides/what-can-i-run-on-16gb-vram/)
- [Quels LLM avec 24 Go de VRAM? (2026)](https://bestlocalairig.com/fr/guides/what-can-i-run-on-24gb-vram/)
- [Quels LLM avec 32 Go de VRAM? (2026)](https://bestlocalairig.com/fr/guides/what-can-i-run-on-32gb-vram/)
- [Quels LLM avec 64 Go de mémoire unifiée? (2026)](https://bestlocalairig.com/fr/guides/what-can-i-run-on-64gb-unified-memory/)
- [Quels LLM avec 96 Go de mémoire unifiée? (2026)](https://bestlocalairig.com/fr/guides/what-can-i-run-on-96gb-unified-memory/)
- [Quels LLM avec 128 Go de mémoire unifiée? (2026)](https://bestlocalairig.com/fr/guides/what-can-i-run-on-128gb-unified-memory/)
- [Meilleure config pour l’IA locale à moins de 2 000 $ US (2026)](https://bestlocalairig.com/fr/guides/best-local-ai-rig-under-2000/)
- [Meilleure config pour l’IA locale à moins de 2 500 $ US (2026)](https://bestlocalairig.com/fr/guides/best-local-ai-rig-under-2500/)
- [Meilleure config pour l’IA locale à moins de 3 500 $ US (2026)](https://bestlocalairig.com/fr/guides/best-local-ai-rig-under-3500/)
- [Meilleure config pour l’IA locale à moins de 5 000 $ US (2026)](https://bestlocalairig.com/fr/guides/best-local-ai-rig-under-5000/)
- [Mac Studio vs Strix Halo pour l’IA locale : vitesse et prix](https://bestlocalairig.com/fr/guides/mac-studio-vs-strix-halo/)
- [DGX Spark vs Strix Halo pour l’IA locale : vitesse et prix](https://bestlocalairig.com/fr/guides/dgx-spark-vs-strix-halo/)
- [DGX Spark vs Mac Studio pour l’IA locale : vitesse et prix](https://bestlocalairig.com/fr/guides/dgx-spark-vs-mac-studio/)
- [DGX Spark et Llama 70B : vitesse et mémoire (2026)](https://bestlocalairig.com/fr/guides/can-dgx-spark-run-llama-70b/)
- [Meilleur mini PC pour l’IA locale à moins de 2 000 $ US (2026)](https://bestlocalairig.com/fr/guides/best-mini-pc-for-local-llms-under-2000/)
- [Quelle carte graphique pour l’IA locale (2026) : VRAM et prix](https://bestlocalairig.com/fr/guides/best-gpu-for-local-llms/)
- [Combien de RAM/VRAM pour une IA en local? (2026)](https://bestlocalairig.com/fr/guides/how-much-memory-to-run-llms-locally/)
- [Faire tourner un LLM 70B en local au meilleur prix (2026)](https://bestlocalairig.com/fr/guides/cheapest-way-to-run-70b-locally/)
- [Strix Halo pour l’IA locale : essais et prix (2026)](https://bestlocalairig.com/fr/guides/is-strix-halo-good-for-local-llms/)
- [DGX Spark vs RTX 5090 pour l’IA locale (2026)](https://bestlocalairig.com/fr/guides/dgx-spark-vs-rtx-5090/)
- [Qwen3.5 9B en local : VRAM nécessaire et matériel (2026)](https://bestlocalairig.com/fr/models/qwen3-5-9b/)
- [Gemma 4 12B en local : VRAM nécessaire et matériel (2026)](https://bestlocalairig.com/fr/models/gemma-4-12b/)
- [GPT-OSS 20B en local : VRAM nécessaire et matériel (2026)](https://bestlocalairig.com/fr/models/gpt-oss-20b/)
- [Gemma 4 26B-A4B en local : VRAM et matériel (2026)](https://bestlocalairig.com/fr/models/gemma-4-26b-a4b/)
- [Qwen3.8 27B en local : VRAM nécessaire et matériel (2026)](https://bestlocalairig.com/fr/models/qwen3-8-27b/)
- [Gemma 3 27B en local : VRAM nécessaire et matériel (2026)](https://bestlocalairig.com/fr/models/gemma-3-27b/)
- [GLM-4.7-Flash 30B-A3B en local : VRAM et matériel (2026)](https://bestlocalairig.com/fr/models/glm-4-7-flash/)
- [Qwen3 30B-A3B en local : VRAM nécessaire et matériel (2026)](https://bestlocalairig.com/fr/models/qwen3-30b-a3b/)
- [Gemma 4 31B en local : VRAM nécessaire et matériel (2026)](https://bestlocalairig.com/fr/models/gemma-4-31b/)
- [Qwen3 32B en local : VRAM nécessaire et matériel (2026)](https://bestlocalairig.com/fr/models/qwen3-32b/)
- [DeepSeek R1 Distill Qwen 32B en local : VRAM nécessaire](https://bestlocalairig.com/fr/models/deepseek-r1-distill-qwen-32b/)
- [Qwen3.6 35B-A3B en local : VRAM et matériel (2026)](https://bestlocalairig.com/fr/models/qwen3-6-35b-a3b/)
- [Llama 3.3 70B en local : VRAM nécessaire et matériel (2026)](https://bestlocalairig.com/fr/models/llama-3-3-70b/)
- [DeepSeek R1 Distill Llama 70B en local : VRAM nécessaire](https://bestlocalairig.com/fr/models/deepseek-r1-distill-llama-70b/)
- [Qwen3-Next 80B-A3B en local : VRAM et matériel (2026)](https://bestlocalairig.com/fr/models/qwen3-next-80b-a3b/)
- [Llama 4 Scout 109B-A17B en local : VRAM et matériel (2026)](https://bestlocalairig.com/fr/models/llama-4-scout/)
- [GPT-OSS 120B en local : VRAM nécessaire et matériel (2026)](https://bestlocalairig.com/fr/models/gpt-oss-120b/)
- [Mistral Small 4 119B-A6B en local : VRAM et matériel (2026)](https://bestlocalairig.com/fr/models/mistral-small-4-119b/)
- [Nemotron 3 Super 120B-A12B en local : VRAM nécessaire](https://bestlocalairig.com/fr/models/nemotron-3-super-120b/)
- [Qwen3.5 122B-A10B en local : VRAM et matériel (2026)](https://bestlocalairig.com/fr/models/qwen3-5-122b-a10b/)
- [Qwen3 235B-A22B en local : VRAM et matériel (2026)](https://bestlocalairig.com/fr/models/qwen3-235b-a22b/)
- [Muse Spark Open 816B-A42B en local : VRAM et matériel (2026)](https://bestlocalairig.com/fr/models/muse-spark-open/)