Machines · 54 machines · 12 categories · data 2026-10-15

What to run it on.

Memory decides which models fit. Bandwidth decides how fast they talk.

2026 prices are inflated by the memory-chip shortage (an RTX 5090 sells for about 3.5× its launch price). All prices are approximate and dated; check before buying.

Before you buy

Buy a computer, or pay for AI?

Paid AI planAbout $20 a month

That is $240 a year. You get the smartest AI. It needs the internet, and your words go to the company. Check today's price.

The computer you have$0

A 16 GB computer runs small models, about 36% as smart as the best paid AI. Try this first. It costs nothing.

A new 128 GB computerFrom about $3,449

About 59% as smart as the best paid AI, today. That price is about 14 years of a $20 plan.

Our advice: start free on the computer you have. Buy a big one only if you need privacy, work offline, use AI all day, or need a new computer anyway.

48 machines can run the 16 GB model list. Cheapest first.

Used server GPUused
Tesla P40 24GB (used)
Memory
24 GB~23 GB for models
Bandwidth
345.6 GB/ssets speed
From
~$277used · as of 2026-10-02
Best that fit · 56 models
1. Qwen3.8 27B33.7
2. Qwen3.5 27B22.9
3. Qwen3.6 27B21.4
See the 24 GB model list
Single-board computer
Raspberry Pi 5 (16GB), CPU only
Memory
16 GB~12 GB for models
Bandwidth
17 GB/ssets speed
From
~$305as of 2026-10-12
Best that fit · 38 models
1. K2 Horizon 7B20.6
2. Gemma 4 12B14.2
3. Qwen3.5 9B13.3
See the 16 GB model list
Used server GPUused
AMD Instinct MI50 32GB (used)
Memory
32 GB~31 GB for models
Bandwidth
1024 GB/ssets speed
From
~$599used · as of 2026-10-02
Best that fit · 75 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 32 GB model list
Intel GPU
Intel Arc Pro B60 (24GB)
Memory
24 GB~23 GB for models
Bandwidth
456 GB/ssets speed
From
~$650as of 2026-10-15
Best that fit · 56 models
1. Qwen3.8 27B33.7
2. Qwen3.5 27B22.9
3. Qwen3.6 27B21.4
See the 24 GB model list
Used server GPUused
Tesla V100 PCIe 32GB (used)
Memory
32 GB~31 GB for models
Bandwidth
n/asets speed
From
~$670used · as of 2026-10-02
Best that fit · 75 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 32 GB model list
NVIDIA GPU
GeForce RTX 5060 Ti 16GB
Memory
16 GB~15 GB for models
Bandwidth
448 GB/ssets speed
From
~$740as of 2026-10-15
Best that fit · 44 models
1. K2 Horizon 7B20.6
2. Gemma 4 12B14.2
3. Qwen3.5 9B13.3
See the 16 GB model list
AMD GPU
Radeon RX 9070 XT
Memory
16 GB~15 GB for models
Bandwidth
640 GB/ssets speed
From
~$760as of 2026-10-15
Best that fit · 44 models
1. K2 Horizon 7B20.6
2. Gemma 4 12B14.2
3. Qwen3.5 9B13.3
See the 16 GB model list
Apple desktop
Mac mini (M6)
Memory
16 GB~12 GB for models
Bandwidth
153 GB/ssets speed
From
~$899as of 2026-10-15
Best that fit · 38 models
1. K2 Horizon 7B20.6
2. Gemma 4 12B14.2
3. Qwen3.5 9B13.3
See the 16 GB model list
AMD GPUprevious gen
Radeon RX 7900 XTX
Memory
24 GB~23 GB for models
Bandwidth
960 GB/ssets speed
From
~$950used · as of 2026-10-02
Best that fit · 56 models
1. Qwen3.8 27B33.7
2. Qwen3.5 27B22.9
3. Qwen3.6 27B21.4
See the 24 GB model list
NVIDIA GPU
GeForce RTX 5070 Ti
Memory
16 GB~15 GB for models
Bandwidth
896 GB/ssets speed
From
~$1,130as of 2026-10-15
Best that fit · 44 models
1. K2 Horizon 7B20.6
2. Gemma 4 12B14.2
3. Qwen3.5 9B13.3
See the 16 GB model list
Intel GPU
Intel Arc Pro B70 (32GB)
Memory
32 GB~31 GB for models
Bandwidth
608 GB/ssets speed
From
~$1,280as of 2026-10-15
Best that fit · 75 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 32 GB model list
Apple laptop
MacBook Air 13"/15" (M5)
Memory
16 GB~12 GB for models
Bandwidth
153 GB/ssets speed
From
~$1,299as of 2026-10-15
Best that fit · 38 models
1. K2 Horizon 7B20.6
2. Gemma 4 12B14.2
3. Qwen3.5 9B13.3
See the 16 GB model list
NVIDIA GPUused
GeForce RTX 3090 (used)
Memory
24 GB~23 GB for models
Bandwidth
936 GB/ssets speed
From
~$1,430used · as of 2026-10-02
Best that fit · 56 models
1. Qwen3.8 27B33.7
2. Qwen3.5 27B22.9
3. Qwen3.6 27B21.4
See the 24 GB model list
NVIDIA GPU
GeForce RTX 5080
Memory
16 GB~15 GB for models
Bandwidth
960 GB/ssets speed
From
~$1,550as of 2026-10-15
Best that fit · 44 models
1. K2 Horizon 7B20.6
2. Gemma 4 12B14.2
3. Qwen3.5 9B13.3
See the 16 GB model list
AMD GPU
Radeon AI PRO R9700 (32GB)
Memory
32 GB~31 GB for models
Bandwidth
640 GB/ssets speed
From
~$1,600as of 2026-10-15
Best that fit · 75 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 32 GB model list
Apple desktop
Mac mini (M5 Pro)
Memory
24 GB~18 GB for models
Bandwidth
307 GB/ssets speed
From
~$1,699as of 2026-10-15
Best that fit · 45 models
1. K2 Horizon 7B20.6
2. Gemma 4 12B14.2
3. Qwen3.5 9B13.3
See the 24 GB model list
AMD mini PC
Framework Desktop (Ryzen AI Max+ 395)
Memory
64 GB~48 GB for models
Bandwidth
256 GB/ssets speed
From
~$1,959as of 2026-10-15
Best that fit · 77 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 64 GB model list
Apple laptop
MacBook Pro 14" (M5)
Memory
16 GB~12 GB for models
Bandwidth
153 GB/ssets speed
From
~$1,999base config, approx.
Best that fit · 38 models
1. K2 Horizon 7B20.6
2. Gemma 4 12B14.2
3. Qwen3.5 9B13.3
See the 16 GB model list
AMD mini PC
GMKtec EVO-X2 (Ryzen AI Max+ 395)
Memory
64 GB~48 GB for models
Bandwidth
256 GB/ssets speed
From
~$2,200as of 2026-10-15
Best that fit · 77 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 64 GB model list
Apple laptop
MacBook Pro 14"/16" (M5 Pro)
Memory
24 GB~18 GB for models
Bandwidth
307 GB/ssets speed
From
~$2,499as of 2026-10-15
Best that fit · 45 models
1. K2 Horizon 7B20.6
2. Gemma 4 12B14.2
3. Qwen3.5 9B13.3
See the 24 GB model list
Apple desktop
Mac Studio (M5 Max)
Memory
36 GB~27 GB for models
Bandwidth
614 GB/ssets speed
From
~$2,499as of 2026-10-15
Best that fit · 67 models
1. Qwen3.8 27B33.7
2. Qwen3.5 27B22.9
3. Qwen3.6 27B21.4
See the 32 GB model list
AMD mini PC
Minisforum MS-S1 MAX (Ryzen AI Max+ 395)
Memory
64 GB~48 GB for models
Bandwidth
256 GB/ssets speed
From
~$2,599as of 2026-10-15
Best that fit · 77 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 64 GB model list
NVIDIA workstation GPU
RTX PRO 4000 Blackwell 24GB
Memory
24 GB~23 GB for models
Bandwidth
672 GB/ssets speed
From
~$2,900as of 2026-10-15
Best that fit · 56 models
1. Qwen3.8 27B33.7
2. Qwen3.5 27B22.9
3. Qwen3.6 27B21.4
See the 24 GB model list
NVIDIA GPUprevious gen
GeForce RTX 4090
Memory
24 GB~23 GB for models
Bandwidth
1008 GB/ssets speed
From
~$3,099used · as of 2026-10-02
Best that fit · 56 models
1. Qwen3.8 27B33.7
2. Qwen3.5 27B22.9
3. Qwen3.6 27B21.4
See the 24 GB model list
AMD mini PC
Bosgame M5 (Ryzen AI Max+ 395)
Memory
128 GB~96 GB for models
Bandwidth
256 GB/ssets speed
From
~$3,599as of 2026-10-15
Best that fit · 90 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 128 GB model list
AMD mini PCannounced
iBUYPOWER Pro AMD AI Halo Mini PC (Ryzen AI Max+ PRO 495)
Memory
128 GB~96 GB for models
Bandwidth
n/asets speed
From
~$3,999as of 2026-10-16
Best that fit · 90 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 128 GB model list
Apple laptop
MacBook Pro 14"/16" (M5 Max)
Memory
36 GB~27 GB for models
Bandwidth
614 GB/ssets speed
From
~$4,099as of 2026-10-15
Best that fit · 67 models
1. Qwen3.8 27B33.7
2. Qwen3.5 27B22.9
3. Qwen3.6 27B21.4
See the 32 GB model list
AMD mini PC
Beelink GTR9 Pro (Ryzen AI Max+ 395)
Memory
128 GB~96 GB for models
Bandwidth
256 GB/ssets speed
From
~$4,349as of 2026-10-15
Best that fit · 90 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 128 GB model list
NVIDIA workstation GPU
RTX PRO 4500 Blackwell 32GB
Memory
32 GB~31 GB for models
Bandwidth
n/asets speed
From
~$4,500as of 2026-10-15
Best that fit · 75 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 32 GB model list
NVIDIA AI boxannounced
DGX Spark 64GB (partner GB10 systems)
Memory
64 GB~48 GB for models
Bandwidth
273 GB/ssets speed
From
~$4,999as of 2026-10-15
Best that fit · 77 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 64 GB model list
Apple desktop
Mac Studio (M5 Ultra)
Memory
96 GB~72 GB for models
Bandwidth
1200 GB/ssets speed
From
~$5,499as of 2026-10-15
Best that fit · 83 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 96 GB model list
Edge dev kit
NVIDIA Jetson AGX Thor Developer Kit
Memory
128 GB~96 GB for models
Bandwidth
273 GB/ssets speed
From
~$5,499as of 2026-10-15
Best that fit · 90 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 128 GB model list
NVIDIA GPU
GeForce RTX 5090
Memory
32 GB~31 GB for models
Bandwidth
1792 GB/ssets speed
From
~$5,700as of 2026-10-15
Best that fit · 75 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 32 GB model list
NVIDIA AI boxannounced
Microsoft Surface RTX Spark Dev Box (128GB)
Memory
128 GB~96 GB for models
Bandwidth
n/asets speed
From
~$6,000as of 2026-10-16
Best that fit · 90 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 128 GB model list
AMD mini PC
GMKtec EVO-X5 Pro (Ryzen AI Max+ PRO 495 "Gorgon Halo")
Memory
192 GB~159.9 GB for models
Bandwidth
273 GB/ssets speed
From
~$6,799as of 2026-10-15
Best that fit · 94 models
1. Qwen3.8-Flash-Next (125B-A6B)39.8
2. Qwen3.8 27B33.7
3. Nemotron 3.5 Super VL 120B-A12B26.5
See the 192 GB model list
NVIDIA AI box
NVIDIA DGX Spark (128GB)
Memory
128 GB~96 GB for models
Bandwidth
273 GB/ssets speed
From
~$6,950as of 2026-10-15
Best that fit · 90 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 128 GB model list
AMD mini PC
Minisforum MS-S1 MAX-P495 (Ryzen AI Max+ PRO 495 "Gorgon Halo")
Memory
192 GB~159.9 GB for models
Bandwidth
n/asets speed
From
~$7,399as of 2026-10-16
Best that fit · 94 models
1. Qwen3.8-Flash-Next (125B-A6B)39.8
2. Qwen3.8 27B33.7
3. Nemotron 3.5 Super VL 120B-A12B26.5
See the 192 GB model list
NVIDIA workstation GPU
RTX PRO 5000 Blackwell 72GB
Memory
72 GB~71 GB for models
Bandwidth
1344 GB/ssets speed
From
~$9,500as of 2026-10-15
Best that fit · 83 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 64 GB model list
NVIDIA workstation GPU
RTX PRO 6000 Blackwell Workstation (96GB)
Memory
96 GB~95 GB for models
Bandwidth
1792 GB/ssets speed
From
~$14,900as of 2026-10-15
Best that fit · 89 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 96 GB model list
NVIDIA workstation GPU
RTX PRO 6000 Blackwell Max-Q (96GB)
Memory
96 GB~95 GB for models
Bandwidth
1792 GB/ssets speed
From
~$14,900as of 2026-10-15
Best that fit · 89 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 96 GB model list
NVIDIA AI box
Gigabyte W775-V10-L01 (DGX Station GB300)
Memory
252 GB~251 GB for models
Bandwidth
7100 GB/ssets speed
From
~$119,000as of 2026-10-16
Best that fit · 97 models
1. GLM-5.3-Flash41.8
2. Qwen3.8-Flash-Next (125B-A6B)39.8
3. MiMo-V2.6-Flash37.9
See the 256 GB model list
NVIDIA GPUprevious gen
GeForce RTX 4080 Super
Memory
16 GB~15 GB for models
Bandwidth
736 GB/ssets speed
From
n/ano current price
Best that fit · 44 models
1. K2 Horizon 7B20.6
2. Gemma 4 12B14.2
3. Qwen3.5 9B13.3
See the 16 GB model list
NVIDIA GPUprevious gen
GeForce RTX 4070 Ti Super
Memory
16 GB~15 GB for models
Bandwidth
672 GB/ssets speed
From
n/ano current price
Best that fit · 44 models
1. K2 Horizon 7B20.6
2. Gemma 4 12B14.2
3. Qwen3.5 9B13.3
See the 16 GB model list
NVIDIA GPUprevious gen
GeForce RTX 4060 Ti 16GB
Memory
16 GB~15 GB for models
Bandwidth
288 GB/ssets speed
From
n/ano current price
Best that fit · 44 models
1. K2 Horizon 7B20.6
2. Gemma 4 12B14.2
3. Qwen3.5 9B13.3
See the 16 GB model list
NVIDIA GPUused
GeForce RTX 3090 Ti
Memory
24 GB~23 GB for models
Bandwidth
1008 GB/ssets speed
From
n/ano current price
Best that fit · 56 models
1. Qwen3.8 27B33.7
2. Qwen3.5 27B22.9
3. Qwen3.6 27B21.4
See the 24 GB model list
NVIDIA AI boxannounced
ASUS ProArt GR1X (NVIDIA RTX Spark)
Memory
64 GB~48 GB for models
Bandwidth
n/asets speed
From
n/ano current price
Best that fit · 77 models
1. Qwen3.8 27B33.7
2. K2 Horizon MoVA 36B-A4B25.3
3. Qwen3.5 27B22.9
See the 64 GB model list
AMD mini PC
AMD Ryzen AI Halo Developer Platform (Ryzen AI Max+ PRO 495, 192GB)
Memory
192 GB~159.9 GB for models
Bandwidth
n/asets speed
From
n/ano current price
Best that fit · 94 models
1. Qwen3.8-Flash-Next (125B-A6B)39.8
2. Qwen3.8 27B33.7
3. Nemotron 3.5 Super VL 120B-A12B26.5
See the 192 GB model list
Edge dev kitannounced
Sipeed MaixStation
Memory
32 GB~24 GB for models
Bandwidth
273 GB/ssets speed
From
n/ano current price
Best that fit · 62 models
1. Qwen3.8 27B33.7
2. Qwen3.5 27B22.9
3. Qwen3.6 27B21.4
See the 32 GB model list

Speed

Same model, different machines.

Published decode speeds for OpenAI's open gpt-oss models.

`gpt-oss-120b`6 published runs · decode tok/s, higher is faster

Mac Studio (M5 Ultra)oMLX v0.7.0 · MXFP4 (MLX)110.1^c
MacBook Pro 14"/16" (M5 Max)Ollama 0.34.1 · MXFP4 (GGUF)79.1^c
Mac Studio (M5 Max) · Ryan's rigoMLX v0.3.9 · MXFP4-Q8 (MLX)78.2^c
NVIDIA DGX Spark (128GB)llama.cpp · MXFP455.37
Framework Desktop (Ryzen AI Max+ 395)llama.cpp Vulkan · MXFP4 (GGUF)54^c
GMKtec EVO-X2 (Ryzen AI Max+ 395)llama.cpp ROCm (llama-bench) · MXFP4 (GGUF)46.05^c

`gpt-oss-20b`10 published runs · decode tok/s, higher is faster

GeForce RTX 5090llama.cpp · MXFP4318.4^c
GeForce RTX 4090llama.cpp · Q4_K_M242.6^c
GeForce RTX 3090 (used)llama.cpp (Linux, flash attention) · MXFP4147.53^c
Radeon RX 9070 XTllama.cpp · unspecified140^?
MacBook Pro 14"/16" (M5 Max)Ollama 0.34.1 · MXFP4 (GGUF)113.4^c
NVIDIA DGX Spark (128GB)llama.cpp · MXFP482.74
MacBook Pro 14"/16" (M5 Pro)oMLX v0.6.2 · MXFP4-Q8 (MLX)76.6^c
Mac mini (M5 Pro)oMLX v0.6.2 · MXFP4-Q8 (MLX)76.6^c
Minisforum MS-S1 MAX (Ryzen AI Max+ 395)llama.cpp ROCm 7.1 · UD-Q4_K_XL75.7^c
MacBook Pro 14" (M5)oMLX v0.3.8 · MXFP4 (MLX)48.7^c

^c community-measured · ^? community, unverified. Published numbers only, from different engines and context lengths, so they aren't strictly comparable. Click a number for its source.

Usable memory: 75% of RAM on Macs (the macOS default), the GPU share on AMD mini PCs, VRAM minus 1 GB on graphics cards. Single card only; the Raspberry Pi AI HAT+ 2 is left out of the picker because it only runs vendor-compiled models. Speeds are copied from published benchmarks, not computed or tested by us.

AGENT VIEW · text/markdown · or send Accept: text/markdown

# Inside the Machine — Local-AI Hardware Directory (machines.json)

Human page: https://insidethemachine.ai/machines · JSON: https://insidethemachine.ai/machines.json · Models by RAM tier: https://insidethemachine.ai/models.md

Generated 2026-10-15 · 47 machines · prices APPROXIMATE, USD, timestamped, linked. 2026 prices are inflated by the DRAM shortage — re-check before quoting.

## JSON schema (machines.json, schema_version 1.0)

```
{
  schema_version, generated, generated_by,
  methodology: {prices, usable_memory, tier_mapping, tokens_per_sec, speed_rule_of_thumb, power},
  sources: [{key, name, url, as_of}],
  ram_tier_map: [{tier_gb, machines:[{id, name, memory_gb, bandwidth_gbps, price_from_usd}]}],
  machines: [{
    id, name, category, vendor, chip, form_factor, status,          // status: current|previous_gen|used|announced
    memory_type,            // unified | vram | system | accelerator_dram
    memory_options_gb: [int], max_memory_gb, memory_tech,
    bandwidth_gbps, bandwidth_note, spec_source:{source_name,url,as_of},
    gpu_usable_fraction?,   // AMD APUs only (GPU-allocatable share)
    launch_msrp_usd,
    prices: [{value_usd, approx:true, basis, config, condition:new|used, retailer, url, as_of}],
    price_from_usd,         // min of prices[]
    power_w: {value, basis, source_name, url, as_of} | null,
    tokens_per_sec: [{model, quant, engine, context_tokens, tg_tok_s, pp_tok_s, config,
                      source_type: vendor-measured|community-measured|community-unverified,
                      source_name, url, as_of, note}],
    configs: [{memory_gb, usable_for_models_gb, usable_basis, equivalent_ram_tier_gb,
               models_that_fit, best_models_fit:[{id,name,aa_index_best,min_memory_gb}]}],
    ram_tiers_supported: [int], notes
  }]
}
```

## Methodology

- **prices**: All prices are APPROXIMATE USD, timestamped (as_of) and linked. 2026 prices are heavily inflated by the DRAM shortage (e.g. RTX 5090 street ≈3.5× MSRP; Strix Halo 128GB ≈+73% vs launch). basis values: starting_at_list (base config list price), list, retail_new, street_new_tracked_listing, lowest_in_stock_new, lowest_listed_new(_OEM), ebay_median_sold_30d (used), ebay_single_sale, derived_from_list. Re-check before quoting.
- **usable_memory**: usable_for_models_gb: unified Apple = 0.75 × RAM (default macOS GPU wired limit; raise with sudo sysctl iogpu.wired_limit_mb); AMD APUs use the GPU-allocatable share from the source; NVIDIA unified (DGX Spark/Jetson) and Pi use 0.75 as an ESTIMATE; discrete GPUs = VRAM − 1 GB (ESTIMATE). Single card only – multi-GPU pooling not modeled.
- **tier_mapping**: equivalent_ram_tier_gb = largest standard tier T (8…512) with 0.75 × T ≤ usable_for_models_gb, i.e. the models.json tier whose fit list this config can run. models_that_fit counts open-weight non-embedding models in models.json with min_memory_gb ≤ usable_for_models_gb.
- **tokens_per_sec**: Only published numbers, each with source/url/date/engine/quant. source_type: vendor-measured, community-measured, community-unverified. Numbers from different engines/contexts are not strictly comparable. tg = decode (generation) tok/s, pp = prompt processing tok/s.
- **speed_rule_of_thumb**: Decode speed scales roughly with bandwidth_gbps ÷ GB of weights read per token, but MoE overheads, MTP/speculative decoding and engine differences make this unreliable as a number, so no computed speeds are included – only published measurements.
- **power**: power_w is vendor TGP/TDP for cards, max continuous for Apple desktops (not typical draw), measured-under-load where the source says so.

Categories: amd_apu_minipc, amd_gpu, apple_desktop, apple_laptop, edge_accelerator, edge_devkit, edge_sbc, intel_gpu, nvidia_ai_box, nvidia_consumer_gpu, nvidia_workstation_gpu, used_server_gpu

## RAM tier → machines (joins to models.json `ram_tiers`)

| Tier | Machines (config) |
|---|---|
| 8GB | GeForce RTX 5070 12GB; GeForce RTX 5060 8GB; GeForce RTX 3060 12GB (used) 12GB; NVIDIA Jetson Orin Nano Super Developer Kit 8GB; Intel Arc B580 12GB |
| 16GB | MacBook Air 13"/15" (M5) 16GB; MacBook Pro 14" (M5) 16GB; Mac mini (M6) 16GB; GeForce RTX 5080 16GB; GeForce RTX 5070 Ti 16GB; GeForce RTX 5060 Ti 16GB 16GB; GeForce RTX 4080 Super 16GB; GeForce RTX 4070 Ti Super 16GB; GeForce RTX 4060 Ti 16GB 16GB; Radeon RX 9070 XT 16GB; Raspberry Pi 5 (16GB), CPU only 16GB |
| 24GB | MacBook Air 13"/15" (M5) 24GB; MacBook Pro 14" (M5) 24GB; MacBook Pro 14"/16" (M5 Pro) 24GB; Mac mini (M6) 24GB; Mac mini (M5 Pro) 24GB; GeForce RTX 4090 24GB; GeForce RTX 3090 (used) 24GB; GeForce RTX 3090 Ti 24GB; RTX PRO 4000 Blackwell 24GB 24GB; Radeon RX 7900 XTX 24GB; Intel Arc Pro B60 (24GB) 24GB; Tesla P40 24GB (used) 24GB |
| 32GB | MacBook Air 13"/15" (M5) 32GB; MacBook Pro 14" (M5) 32GB; MacBook Pro 14"/16" (M5 Max) 36GB; Mac mini (M6) 32GB; Mac Studio (M5 Max) 36GB; GeForce RTX 5090 32GB; RTX PRO 4500 Blackwell 32GB 32GB; Radeon AI PRO R9700 (32GB) 32GB; Intel Arc Pro B70 (32GB) 32GB; AMD Instinct MI50 32GB (used) 32GB; Tesla V100 PCIe 32GB (used) 32GB; Sipeed MaixStation 32GB |
| 48GB | MacBook Pro 14"/16" (M5 Pro) 48GB; MacBook Pro 14"/16" (M5 Max) 48GB; Mac mini (M5 Pro) 48GB; Mac Studio (M5 Max) 48GB |
| 64GB | MacBook Pro 14"/16" (M5 Pro) 64GB; MacBook Pro 14"/16" (M5 Max) 64GB; Mac mini (M5 Pro) 64GB; Mac Studio (M5 Max) 64GB; RTX PRO 5000 Blackwell 72GB 72GB; DGX Spark 64GB (partner GB10 systems) 64GB; Framework Desktop (Ryzen AI Max+ 395) 64GB; GMKtec EVO-X2 (Ryzen AI Max+ 395) 64GB; Minisforum MS-S1 MAX (Ryzen AI Max+ 395) 64GB; ASUS ProArt GR1X (NVIDIA RTX Spark) 64GB |
| 96GB | Mac Studio (M5 Ultra) 96GB; RTX PRO 6000 Blackwell Workstation (96GB) 96GB; RTX PRO 6000 Blackwell Max-Q (96GB) 96GB |
| 128GB | MacBook Pro 14"/16" (M5 Max) 128GB; Mac Studio (M5 Max) 128GB; NVIDIA DGX Spark (128GB) 128GB; NVIDIA Jetson AGX Thor Developer Kit 128GB; Framework Desktop (Ryzen AI Max+ 395) 128GB; GMKtec EVO-X2 (Ryzen AI Max+ 395) 128GB; Minisforum MS-S1 MAX (Ryzen AI Max+ 395) 128GB; Bosgame M5 (Ryzen AI Max+ 395) 128GB; Beelink GTR9 Pro (Ryzen AI Max+ 395) 128GB; ASUS ProArt GR1X (NVIDIA RTX Spark) 128GB; iBUYPOWER Pro AMD AI Halo Mini PC (Ryzen AI Max+ PRO 495) 128GB; Microsoft Surface RTX Spark Dev Box (128GB) 128GB; Sipeed MaixStation 128GB |
| 192GB | GMKtec EVO-X5 Pro (Ryzen AI Max+ PRO 495 "Gorgon Halo") 192GB; Minisforum MS-S1 MAX-P495 (Ryzen AI Max+ PRO 495 "Gorgon Halo") 192GB; AMD Ryzen AI Halo Developer Platform (Ryzen AI Max+ PRO 495, 192GB) 192GB |
| 256GB | Mac Studio (M5 Ultra) 256GB; Gigabyte W775-V10-L01 (DGX Station GB300) 252GB |
| 512GB | Mac Studio (M5 Ultra) 512GB |

## All machines

| Machine | Category | Memory options (GB) | Bandwidth GB/s | From (USD, approx) | Price basis / date | Power W | Tiers |
|---|---|---|---|---|---|---|---|
| MacBook Air 13"/15" (M5) | apple_laptop | 16, 24, 32 | 153 | [$1,299](https://www.apple.com/shop/buy-mac/macbook-air) | starting_at_list, new, 2026-10-15 | — | 16, 24, 32 |
| MacBook Pro 14" (M5) | apple_laptop | 16, 24, 32 | 153 | [$1,999](https://www.apple.com/shop/buy-mac/macbook-pro) | starting_at_list, new, 2026-10-15 | — | 16, 24, 32 |
| MacBook Pro 14"/16" (M5 Pro) | apple_laptop | 24, 48, 64 | 307 | [$2,499](https://www.apple.com/shop/buy-mac/macbook-pro) | starting_at_list, new, 2026-10-15 | — | 24, 48, 64 |
| MacBook Pro 14"/16" (M5 Max) | apple_laptop | 36, 48, 64, 128 | 614 | [$4,099](https://www.apple.com/shop/buy-mac/macbook-pro) | starting_at_list, new, 2026-10-15 | — | 32, 48, 64, 128 |
| Mac mini (M6) | apple_desktop | 16, 24, 32 | 153 | [$899](https://www.apple.com/shop/buy-mac/mac-mini) | starting_at_list, new, 2026-10-15 | 155 | 16, 24, 32 |
| Mac mini (M5 Pro) | apple_desktop | 24, 48, 64 | 307 | [$1,699](https://www.apple.com/shop/buy-mac/mac-mini) | starting_at_list, new, 2026-10-15 | 155 | 24, 48, 64 |
| Mac Studio (M5 Max) | apple_desktop | 36, 48, 64, 128 | 614 | [$2,499](https://www.apple.com/shop/buy-mac/mac-studio) | starting_at_list, new, 2026-10-15 | 480 | 32, 48, 64, 128 |
| Mac Studio (M5 Ultra) | apple_desktop | 96, 256, 512 | 1200 | [$5,499](https://www.apple.com/shop/buy-mac/mac-studio) | starting_at_list, new, 2026-10-15 | 480 | 96, 256, 512 |
| GeForce RTX 5090 | nvidia_consumer_gpu | 32 | 1792 | [$5,700](https://rtx.live/) | lowest_in_stock_new, new, 2026-10-15 | 575 | 32 |
| GeForce RTX 5080 | nvidia_consumer_gpu | 16 | 960 | [$1,550](https://videocardprices.com/gpu-price-index.csv) | street_new_tracked_listing, new, 2026-10-15 | 360 | 16 |
| GeForce RTX 5070 Ti | nvidia_consumer_gpu | 16 | 896 | [$1,130](https://rtx.live/) | lowest_in_stock_new, new, 2026-10-15 | 300 | 16 |
| GeForce RTX 5070 | nvidia_consumer_gpu | 12 | 672 | [$720](https://videocardprices.com/gpu-price-index.csv) | street_new_tracked_listing, new, 2026-10-15 | 250 | 8 |
| GeForce RTX 5060 Ti 16GB | nvidia_consumer_gpu | 16 | 448 | [$740](https://rtx.live/) | lowest_in_stock_new, new, 2026-10-15 | 180 | 16 |
| GeForce RTX 5060 | nvidia_consumer_gpu | 8 | 448 | [$370](https://rtx.live/) | lowest_in_stock_new, new, 2026-10-15 | 145 | 8 |
| GeForce RTX 4090 | nvidia_consumer_gpu | 24 | 1008 | [$3,099](https://sillectus.com/media/budget-ai-server-build/ebay-sold-gpu-prices-2026-10-02.csv) | ebay_single_sale, used, 2026-10-02 | 450 | 24 |
| GeForce RTX 4080 Super | nvidia_consumer_gpu | 16 | 736 | — | not captured | 320 | 16 |
| GeForce RTX 4070 Ti Super | nvidia_consumer_gpu | 16 | 672 | — | not captured | 285 | 16 |
| GeForce RTX 4060 Ti 16GB | nvidia_consumer_gpu | 16 | 288 | — | not captured | 165 | 16 |
| GeForce RTX 3090 (used) | nvidia_consumer_gpu | 24 | 936 | [$1,430](https://sillectus.com/media/budget-ai-server-build/ebay-sold-gpu-prices-2026-10-02.csv) | ebay_median_sold_30d, used, 2026-10-02 | 350 | 24 |
| GeForce RTX 3090 Ti | nvidia_consumer_gpu | 24 | 1008 | — | not captured | 450 | 24 |
| GeForce RTX 3060 12GB (used) | nvidia_consumer_gpu | 12 | 360 | [$300](https://sillectus.com/media/budget-ai-server-build/ebay-sold-gpu-prices-2026-10-02.csv) | ebay_median_sold_30d, used, 2026-10-02 | 170 | 8 |
| RTX PRO 6000 Blackwell Workstation (96GB) | nvidia_workstation_gpu | 96 | 1792 | [$14,900](https://vga.live/cards/workstation) | lowest_listed_new_OEM, new, 2026-10-15 | — | 96 |
| RTX PRO 6000 Blackwell Max-Q (96GB) | nvidia_workstation_gpu | 96 | 1792 | [$14,900](https://vga.live/cards/workstation) | lowest_listed_new_OEM, new, 2026-10-15 | — | 96 |
| RTX PRO 5000 Blackwell 72GB | nvidia_workstation_gpu | 72 | 1344 | [$9,500](https://vga.live/cards/workstation) | lowest_listed_new, new, 2026-10-15 | — | 64 |
| RTX PRO 4500 Blackwell 32GB | nvidia_workstation_gpu | 32 | — | [$4,500](https://vga.live/cards/workstation) | lowest_listed_new_OEM, new, 2026-10-15 | — | 32 |
| RTX PRO 4000 Blackwell 24GB | nvidia_workstation_gpu | 24 | 672 | [$2,900](https://vga.live/cards/workstation) | lowest_listed_new_OEM, new, 2026-10-15 | — | 24 |
| NVIDIA DGX Spark (128GB) | nvidia_ai_box | 128 | 273 | [$6,950](https://marketplace.nvidia.com/en-us/enterprise/personal-ai-supercomputers/dgx-spark/) | list, new, 2026-10-15 | 240 | 128 |
| DGX Spark 64GB (partner GB10 systems) | nvidia_ai_box | 64 | 273 | [$4,999](https://blogs.nvidia.com/blog/local-ai-dgx-spark-64gb-sync/) | starting_at_announced, new, 2026-10-15 | — | 64 |
| NVIDIA Jetson AGX Thor Developer Kit | edge_devkit | 128 | 273 | [$5,499](https://www.microcenter.com/product/711571/nvidia-jetson-agx-thor-developer-kit) | retail_new, new, 2026-10-15 | 130 | 128 |
| NVIDIA Jetson Orin Nano Super Developer Kit | edge_devkit | 8 | 102 | [$399](https://computingforgeeks.com/edge-llm-jetson-orin-vs-ai-hat-2-vs-hailo-10h/) | list, new, 2026-10-12 | 25 | 8 |
| Framework Desktop (Ryzen AI Max+ 395) | amd_apu_minipc | 64, 128 | 256 | [$1,959](https://computingforgeeks.com/ryzen-ai-max-395-mini-pc-comparison/) | list, new, 2026-10-15 | 133 | 64, 128 |
| GMKtec EVO-X2 (Ryzen AI Max+ 395) | amd_apu_minipc | 64, 128 | 256 | [$2,200](https://computingforgeeks.com/ryzen-ai-max-395-mini-pc-comparison/) | list, new, 2026-10-15 | 150 | 64, 128 |
| Minisforum MS-S1 MAX (Ryzen AI Max+ 395) | amd_apu_minipc | 64, 128 | 256 | [$2,599](https://computingforgeeks.com/ryzen-ai-max-395-mini-pc-comparison/) | list, new, 2026-10-15 | — | 64, 128 |
| Bosgame M5 (Ryzen AI Max+ 395) | amd_apu_minipc | 128 | 256 | [$3,599](https://computingforgeeks.com/ryzen-ai-max-395-mini-pc-comparison/) | list, new, 2026-10-15 | — | 128 |
| Beelink GTR9 Pro (Ryzen AI Max+ 395) | amd_apu_minipc | 128 | 256 | [$4,349](https://computingforgeeks.com/ryzen-ai-max-395-mini-pc-comparison/) | list, new, 2026-10-15 | — | 128 |
| GMKtec EVO-X5 Pro (Ryzen AI Max+ PRO 495 "Gorgon Halo") | amd_apu_minipc | 192 | 273 | [$6,799](https://localboxes.org/boxes/gmktec-evo-x5-pro) | list, new, 2026-10-15 | — | 192 |
| Radeon RX 7900 XTX | amd_gpu | 24 | 960 | [$950](https://sillectus.com/media/budget-ai-server-build/ebay-sold-gpu-prices-2026-10-02.csv) | ebay_median_sold_30d, used, 2026-10-02 | 355 | 24 |
| Radeon RX 9070 XT | amd_gpu | 16 | 640 | [$760](https://videocardprices.com/gpu-price-index.csv) | street_new_tracked_listing, new, 2026-10-15 | 304 | 16 |
| Radeon AI PRO R9700 (32GB) | amd_gpu | 32 | 640 | [$1,600](https://vga.live/cards/workstation) | lowest_listed_new, new, 2026-10-15 | — | 32 |
| Intel Arc B580 | intel_gpu | 12 | 456 | [$320](https://videocardprices.com/gpu-price-index.csv) | street_new_tracked_listing, new, 2026-10-15 | — | 8 |
| Intel Arc Pro B60 (24GB) | intel_gpu | 24 | 456 | [$650](https://vga.live/cards/workstation) | lowest_listed_new, new, 2026-10-15 | 200 | 24 |
| Intel Arc Pro B70 (32GB) | intel_gpu | 32 | 608 | [$1,280](https://vga.live/cards/workstation) | lowest_listed_new, new, 2026-10-15 | 230 | 32 |
| Raspberry Pi 5 (16GB), CPU only | edge_sbc | 16 | 17 | [$305](https://computingforgeeks.com/edge-llm-jetson-orin-vs-ai-hat-2-vs-hailo-10h/) | list, new, 2026-10-12 | 12.3 | 16 |
| Raspberry Pi AI HAT+ 2 (Hailo-10H, 8GB) on Pi 5 | edge_accelerator | 8 | — | [$200](https://computingforgeeks.com/edge-llm-jetson-orin-vs-ai-hat-2-vs-hailo-10h/) | list, new, 2026-10-12 | 7.1 |  |
| Tesla P40 24GB (used) | used_server_gpu | 24 | 345.6 | [$277](https://sillectus.com/media/budget-ai-server-build/ebay-sold-gpu-prices-2026-10-02.csv) | ebay_median_sold_30d, used, 2026-10-02 | — | 24 |
| AMD Instinct MI50 32GB (used) | used_server_gpu | 32 | 1024 | [$599](https://sillectus.com/media/budget-ai-server-build/ebay-sold-gpu-prices-2026-10-02.csv) | ebay_median_sold_30d, used, 2026-10-02 | 300 | 32 |
| Tesla V100 PCIe 32GB (used) | used_server_gpu | 32 | — | [$670](https://sillectus.com/media/budget-ai-server-build/ebay-sold-gpu-prices-2026-10-02.csv) | ebay_median_sold_30d, used, 2026-10-02 | — | 32 |
| Minisforum MS-S1 MAX-P495 (Ryzen AI Max+ PRO 495 "Gorgon Halo") | amd_apu_minipc | 192 | — | [$7,399](https://store.minisforum.com/products/minisforum-ms-s1-max-p495-ai-workstation) | retail_new, new, 2026-10-16 | 320 | 192 |
| Gigabyte W775-V10-L01 (DGX Station GB300) | nvidia_ai_box | 252 | 7100 | [$119,000](https://www.disctech.com/Gigabyte-W775-V10-L01-AI-Workstation-NVIDIA-GB300-Grace-72-Core-CPU-496GB-LPDDR5X-RAM-Blackwell-GPU-) | retail_new, new, 2026-10-16 | 1600 | 256 |
| ASUS ProArt GR1X (NVIDIA RTX Spark) | nvidia_ai_box | 64, 128 | — | — | not captured | 240 | 64, 128 |
| iBUYPOWER Pro AMD AI Halo Mini PC (Ryzen AI Max+ PRO 495) | amd_apu_minipc | 128 | — | [$3,999](https://www.ibuypower.com/blog/news-pr/ryzen-ai-max-pro) | starting_at_announced, new, 2026-10-16 | — | 128 |
| Microsoft Surface RTX Spark Dev Box (128GB) | nvidia_ai_box | 128 | — | [$6,000](https://www.microsoft.com/en-us/surface/devices/surface-rtx-spark-dev-box) | list, new, 2026-10-16 | — | 128 |
| AMD Ryzen AI Halo Developer Platform (Ryzen AI Max+ PRO 495, 192GB) | amd_apu_minipc | 192 | — | — | not captured | 120 | 192 |
| Sipeed MaixStation | edge_devkit | 32, 128 | 273 | — | not captured | 40 | 32, 128 |

## Published speeds (decode tok/s; not strictly comparable across engines)

| Machine | Model | Quant | Engine | tg tok/s | pp tok/s | Type | Source |
|---|---|---|---|---|---|---|---|
| MacBook Pro 14" (M5) | gpt-oss-20b | MXFP4 (MLX) | oMLX v0.3.8 | 48.7 | 1068 | community-measured | [oMLX benchmark – gpt-oss-20b on M5 (10c)…](https://omlx.ai/benchmarks/performance/xns0rgnb) (2026-10-15) |
| MacBook Pro 14"/16" (M5 Pro) | gpt-oss-20b | MXFP4-Q8 (MLX) | oMLX v0.6.2 | 76.6 | 1611 | community-measured | [oMLX benchmark – gpt-oss-20b MXFP4-Q8 on…](https://omlx.ai/benchmarks/performance/h03e0nvx) (2026-10-15) |
| MacBook Pro 14"/16" (M5 Max) | gpt-oss-120b | MXFP4 (GGUF) | Ollama 0.34.1 | 79.1 | — | community-measured | [llamaperf – gpt-oss community reports (i…](https://llamaperf.com/model/gpt-oss) (2026-10-15) |
| MacBook Pro 14"/16" (M5 Max) | gpt-oss-20b | MXFP4 (GGUF) | Ollama 0.34.1 | 113.4 | — | community-measured | [llamaperf – gpt-oss community reports (i…](https://llamaperf.com/model/gpt-oss) (2026-10-15) |
| MacBook Pro 14"/16" (M5 Max) | Qwen3.8 27B | GGUF Q4 (17 GB) | Ollama 0.34.1 | 36.6 | — | community-measured | [llamaperf – gpt-oss community reports (i…](https://llamaperf.com/model/gpt-oss) (2026-10-15) |
| MacBook Pro 14"/16" (M5 Max) | Qwen3.8 27B | MLX (18 GB) | Ollama 0.34.1 (MLX) | 68.0 | — | community-measured | [llamaperf – gpt-oss community reports (i…](https://llamaperf.com/model/gpt-oss) (2026-10-15) |
| MacBook Pro 14"/16" (M5 Max) | Qwen3.6 35B-A3B | MLX (23 GB) | Ollama 0.34.1 (MLX) | 167.2 | — | community-measured | [llamaperf – gpt-oss community reports (i…](https://llamaperf.com/model/gpt-oss) (2026-10-15) |
| MacBook Pro 14"/16" (M5 Max) | Gemma 4 12B | GGUF Q4 (7.6 GB) | Ollama 0.34.1 | 58.1 | — | community-measured | [llamaperf – gpt-oss community reports (i…](https://llamaperf.com/model/gpt-oss) (2026-10-15) |
| Mac mini (M5 Pro) | gpt-oss-20b | MXFP4-Q8 (MLX) | oMLX v0.6.2 | 76.6 | 1611 | community-measured | [oMLX benchmark – gpt-oss-20b MXFP4-Q8 on…](https://omlx.ai/benchmarks/performance/h03e0nvx) (2026-10-15) |
| Mac Studio (M5 Max) | gpt-oss-120b | MXFP4-Q8 (MLX) | oMLX v0.3.9 | 78.2 | 727.4 | community-measured | [oMLX benchmark – gpt-oss-120b MXFP4-Q8 o…](https://omlx.ai/benchmarks/dgadv0k9) (2026-10-15) |
| Mac Studio (M5 Max) | gpt-oss-120b | MXFP4-Q8 (MLX) | oMLX v0.3.9 | 70.5 | 1576 | community-measured | [oMLX benchmark – gpt-oss-120b MXFP4-Q8 o…](https://omlx.ai/benchmarks/dgadv0k9) (2026-10-15) |
| Mac Studio (M5 Ultra) | gpt-oss-120b | MXFP4 (MLX) | oMLX v0.7.0 | 108.6 | 2829 | community-measured | [oMLX benchmark – gpt-oss-120b on M5 Ultr…](https://omlx.ai/benchmarks/performance/bcoqxlcv) (2026-10-15) |
| Mac Studio (M5 Ultra) | gpt-oss-120b | MXFP4 (MLX) | oMLX v0.7.0 | 110.1 | 1049 | community-measured | [oMLX benchmark – gpt-oss-120b on M5 Ultr…](https://omlx.ai/benchmarks/performance/bcoqxlcv) (2026-10-15) |
| GeForce RTX 5090 | gpt-oss-20b | MXFP4 | llama.cpp | 318.4 | — | community-measured | [llm-speed – gpt-oss-20b signed runs (RTX…](https://llm-speed.com/m/gpt-oss-20b) (2026-10-15) |
| GeForce RTX 4090 | gpt-oss-20b | Q4_K_M | llama.cpp | 242.6 | 593.8 | community-measured | [llm-speed – gpt-oss-20b signed runs (RTX…](https://llm-speed.com/m/gpt-oss-20b) (2026-10-15) |
| GeForce RTX 3090 (used) | gpt-oss-20b | MXFP4 | llama.cpp (Linux, flash attention) | 147.53 | 4400.3 | community-measured | [Hardware Corner – gpt-oss-20b on RTX 309…](https://www.hardware-corner.net/guides/gpt-oss-20b-gpu-benchamrks/) (2025-10-26) |
| NVIDIA DGX Spark (128GB) | gpt-oss-120b | MXFP4 | llama.cpp | 55.37 | 1725.47 | vendor-measured | [NVIDIA Technical Blog – DGX Spark perfor…](https://developer.nvidia.com/blog/how-nvidia-dgx-sparks-performance-enables-intensive-ai-tasks/) (2025-10-24) |
| NVIDIA DGX Spark (128GB) | gpt-oss-20b | MXFP4 | llama.cpp | 82.74 | 3670.42 | vendor-measured | [NVIDIA Technical Blog – DGX Spark perfor…](https://developer.nvidia.com/blog/how-nvidia-dgx-sparks-performance-enables-intensive-ai-tasks/) (2025-10-24) |
| NVIDIA Jetson Orin Nano Super Developer Kit | Llama 3.2 3B | INT4 | MLC | 43.07 | — | vendor-measured | [ComputingForGeeks – Jetson Orin Nano Sup…](https://computingforgeeks.com/edge-llm-jetson-orin-vs-ai-hat-2-vs-hailo-10h/) (2026-10-12) |
| NVIDIA Jetson Orin Nano Super Developer Kit | Qwen2.5 3B | Q4 (Ollama default) | Ollama | 22.6 | — | community-measured | [ComputingForGeeks – Jetson Orin Nano Sup…](https://computingforgeeks.com/edge-llm-jetson-orin-vs-ai-hat-2-vs-hailo-10h/) (2026-10-12) |
| Framework Desktop (Ryzen AI Max+ 395) | gpt-oss-120b | MXFP4 (GGUF) | llama.cpp Vulkan | 54.0 | 500 | community-measured | [llamaperf – gpt-oss community reports (i…](https://llamaperf.com/model/gpt-oss) (2026-10-15) |
| GMKtec EVO-X2 (Ryzen AI Max+ 395) | gpt-oss-120b | MXFP4 (GGUF) | llama.cpp ROCm (llama-bench) | 46.05 | 818.11 | community-measured | [llamaperf – gpt-oss community reports (i…](https://llamaperf.com/model/gpt-oss) (2026-10-15) |
| Minisforum MS-S1 MAX (Ryzen AI Max+ 395) | gpt-oss-20b | UD-Q4_K_XL | llama.cpp ROCm 7.1 | 75.7 | 537.7 | community-measured | [llamaperf – gpt-oss community reports (i…](https://llamaperf.com/model/gpt-oss) (2026-10-15) |
| Radeon RX 9070 XT | gpt-oss-20b | unspecified | llama.cpp | 140 | — | community-unverified | [llm-speed – gpt-oss-20b signed runs (RTX…](https://llm-speed.com/m/gpt-oss-20b) (2026-10-15) |
| Raspberry Pi 5 (16GB), CPU only | llama3.2 3B | Q4 (Ollama default) | Ollama | 5.55 | — | community-measured | [ComputingForGeeks – Jetson Orin Nano Sup…](https://computingforgeeks.com/edge-llm-jetson-orin-vs-ai-hat-2-vs-hailo-10h/) (2026-10-12) |
| Raspberry Pi 5 (16GB), CPU only | qwen2 1.5B | Q4 (Ollama default) | Ollama | 12.47 | — | community-measured | [ComputingForGeeks – Jetson Orin Nano Sup…](https://computingforgeeks.com/edge-llm-jetson-orin-vs-ai-hat-2-vs-hailo-10h/) (2026-10-12) |
| Raspberry Pi AI HAT+ 2 (Hailo-10H, 8GB) on Pi 5 | llama3.2 3B | Hailo HEF | hailo-ollama | 2.66 | — | community-measured | [ComputingForGeeks – Jetson Orin Nano Sup…](https://computingforgeeks.com/edge-llm-jetson-orin-vs-ai-hat-2-vs-hailo-10h/) (2026-10-12) |
| Raspberry Pi AI HAT+ 2 (Hailo-10H, 8GB) on Pi 5 | qwen2 1.5B | Hailo HEF | hailo-ollama | 8.25 | — | community-measured | [ComputingForGeeks – Jetson Orin Nano Sup…](https://computingforgeeks.com/edge-llm-jetson-orin-vs-ai-hat-2-vs-hailo-10h/) (2026-10-12) |

### gpt-oss-120b quick comparison (single machine, published)

| Machine | tg tok/s | Engine | Source |
|---|---|---|---|
| Mac Studio (M5 Ultra) | 110.1 | oMLX v0.7.0 | [2026-10-15](https://omlx.ai/benchmarks/performance/bcoqxlcv) |
| Mac Studio (M5 Ultra) | 108.6 | oMLX v0.7.0 | [2026-10-15](https://omlx.ai/benchmarks/performance/bcoqxlcv) |
| MacBook Pro 14"/16" (M5 Max) | 79.1 | Ollama 0.34.1 | [2026-10-15](https://llamaperf.com/model/gpt-oss) |
| Mac Studio (M5 Max) | 78.2 | oMLX v0.3.9 | [2026-10-15](https://omlx.ai/benchmarks/dgadv0k9) |
| Mac Studio (M5 Max) | 70.5 | oMLX v0.3.9 | [2026-10-15](https://omlx.ai/benchmarks/dgadv0k9) |
| NVIDIA DGX Spark (128GB) | 55.37 | llama.cpp | [2025-10-24](https://developer.nvidia.com/blog/how-nvidia-dgx-sparks-performance-enables-intensive-ai-tasks/) |
| Framework Desktop (Ryzen AI Max+ 395) | 54.0 | llama.cpp Vulkan | [2026-10-15](https://llamaperf.com/model/gpt-oss) |
| GMKtec EVO-X2 (Ryzen AI Max+ 395) | 46.05 | llama.cpp ROCm (llama-bench) | [2026-10-15](https://llamaperf.com/model/gpt-oss) |

## Ryan's machine: Mac Studio M5 Max 128GB

- 614 GB/s; usable ≈ 96.0GB at default wired limit → tier 128GB; 82 catalog models fit.
- Price ≈ $5,099 (128GB/512GB, derived from Apple list $3,099 + $2,000 upgrade per Macworld) / $5,399 at B&H (128GB/1TB), 2026-10-15.
- Measured: gpt-oss-120b 78.2 tok/s (oMLX, 1K ctx) / 70.5 (4K ctx).

## Data gaps

- Apple: exact configured prices for most memory upgrades (only base 'starting at' prices + the 128GB Mac Studio); M5 Ultra 256/512GB configured prices.
- No published tok/s found for: Mac mini M6, MacBook Air M5, RTX 5080/5070 Ti/5070/5060 Ti, RTX 40-series except 4090, RTX PRO cards, R9700, Arc B580/B60/B70, P40, MI50, V100, DGX Spark 64GB, Jetson Thor, Gorgon Halo (device-specific).
- Prices not captured: RTX 4080 Super, 4070 Ti Super, 4060 Ti 16GB, 3090 Ti; RTX 4090 used price is a single sale.
- Power not captured from a fetched source: RTX PRO 6000/5000/4500/4000, Tesla P40, V100, R9700, Arc B580, laptops, most Strix Halo boxes. Apple values are max continuous (PSU), not LLM draw.
- Bandwidth missing: RTX PRO 4500, Tesla V100 32GB, Hailo-10H. Arc B580 bandwidth inferred from the B60 (same die/bus).
- Usable-memory rules for VRAM (−1GB) and NVIDIA/Pi unified (0.75×) are estimates. Multi-GPU builds (2×3090, 4×B60, P40 stacks) and clustering (2× DGX Spark, Strix Halo RPC) are not modeled.
- Not covered yet: laptops with RTX 50 mobile GPUs, Qualcomm Snapdragon X2 / Intel Lunar/Panther Lake NPUs, Tenstorrent, used A100/H100, cloud rental prices.

## Sources

- [Apple – MacBook Air tech specs](https://www.apple.com/macbook-air/specs/) — 2026-10-15
- [Apple Store – Buy MacBook Air](https://www.apple.com/shop/buy-mac/macbook-air) — 2026-10-15
- [Apple – MacBook Pro tech specs](https://www.apple.com/macbook-pro/specs/) — 2026-10-15
- [Apple Store – Buy MacBook Pro](https://www.apple.com/shop/buy-mac/macbook-pro) — 2026-10-15
- [Apple – Mac mini tech specs](https://www.apple.com/mac-mini/specs/) — 2026-10-15
- [Apple Store – Buy Mac mini](https://www.apple.com/shop/buy-mac/mac-mini) — 2026-10-15
- [Apple – Mac Studio tech specs](https://www.apple.com/mac-studio/specs/) — 2026-10-15
- [Apple Store – Buy Mac Studio](https://www.apple.com/shop/buy-mac/mac-studio) — 2026-10-15
- [B&H – Mac Studio M5 Max 40c GPU 128GB/1TB (Z1U4000EF)](https://www.bhphotovideo.com/c/product/1997808-REG/apple_ms_m5m_16_mac_studio_m5_max.html) — 2026-10-15
- [Macworld – Mac Studio (M5 Max) review (upgrade pricing)](https://www.macworld.com/article/3238319/mac-studio-m5-max-review.html) — 2026-10-15
- [NVIDIA – Compare GeForce graphics cards (memory, bandwidth, TGP)](https://www.nvidia.com/en-us/geforce/graphics-cards/compare/) — 2026-10-15
- [Wikipedia – List of Nvidia GPUs (spec tables)](https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units) — 2026-10-15
- [Wikipedia – List of AMD GPUs (spec tables)](https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units) — 2026-10-15
- [videocardprices.com GPU price index (street price, one tracked listing per card)](https://videocardprices.com/gpu-price-index.csv) — 2026-10-15
- [rtx.live – lowest in-stock new price across Micro Center/Best Buy/Newegg/Amazon/etc.](https://rtx.live/) — 2026-10-15
- [gpuprix – RTX 5060 Ti 16GB US price (Newegg/Best Buy/B&H)](https://gpuprix.com/us/gpus/geforce-rtx-5060-ti-16gb) — 2026-10-14
- [vga.live – Workstation GPU prices (Micro Center, Newegg, Amazon, B&H…)](https://vga.live/cards/workstation) — 2026-10-15
- [eBay US sold listings (used), median of last 30 days – sillectus.com dataset](https://sillectus.com/media/budget-ai-server-build/ebay-sold-gpu-prices-2026-10-02.csv) — 2026-10-02
- [VRAMGlass – Arc Pro B60 specs & eBay prices](https://vramglass.com/gpu/arc-pro-b60) — 2026-10-15
- [AMD – Radeon AI PRO R9700 product page](https://www.amd.com/en/products/graphics/workstations/radeon-ai-pro/ai-9000-series/amd-radeon-ai-pro-r9700.html) — 2026-10-15
- [NVIDIA Marketplace – DGX Spark (128GB)](https://marketplace.nvidia.com/en-us/enterprise/personal-ai-supercomputers/dgx-spark/) — 2026-10-15
- [The Register – $4,999 DGX Spark 64GB; 128GB now $6,950 (2 Oct 2026)](https://www.theregister.com/systems/2026/10/02/nvidia-debuts-4999-dgx-spark-with-half-the-ram-and-storage-amid-memory-crunch/5300622) — 2026-10-02
- [NVIDIA Blog – DGX Spark 64GB (partners, from Oct 23, $4,999)](https://blogs.nvidia.com/blog/local-ai-dgx-spark-64gb-sync/) — 2026-10-15
- [NVIDIA Technical Blog – DGX Spark performance (Table 4, ISL 2048 / OSL 128, BS=1)](https://developer.nvidia.com/blog/how-nvidia-dgx-sparks-performance-enables-intensive-ai-tasks/) — 2025-10-24
- [GPUwerk – DGX Spark specs (240W max)](https://gpuwerk.com/dgx-spark/) — 2026-10-15
- [ComputingForGeeks – Ryzen AI Max+ 395 mini PCs compared by price (Oct 2026)](https://computingforgeeks.com/ryzen-ai-max-395-mini-pc-comparison/) — 2026-10-15
- [OpenClaw DC – Is a Strix Halo mini PC worth $3,500 (prices 2026-09-23; 256 GB/s arithmetic)](https://openclawdc.com/blog/is-a-strix-halo-mini-pc-worth-it/) — 2026-09-23
- [sunkcost.ai – Framework Desktop vs EVO-X2 (96GB GPU-usable, power under load)](https://sunkcost.ai/compare/strix-halo-framework-desktop-128gb-vs-strix-halo-gmktec-evo-x2-128gb/) — 2026-10-15
- [Tom's Hardware – GMKtec EVO-X5 (Ryzen AI Max+ PRO 495, 192GB)](https://www.tomshardware.com/desktops/mini-pcs/early-amd-gorgon-halo-ai-mini-pc-packs-192gb-ram-for-an-eye-watering-usd7-099-super-early-bird-deal-cuts-down-price-of-gmktec-evo-x5-with-ryzen-ai-max-pro-495-by-usd425) — 2026-10-15
- [AMD Blog – Meet the next AMD Ryzen AI Halo](https://www.amd.com/en/blogs/2026/meet-the-next-amd-ryzen-ai-halo.html) — 2026-10-15
- [Local Boxes – GMKtec EVO-X5 Pro pricing](https://localboxes.org/boxes/gmktec-evo-x5-pro) — 2026-10-15
- [TechSpot – first Ryzen AI Max+ Pro 495 PCs ($7,399, 192GB)](https://www.techspot.com/news/114021-one-first-ryzen-ai-max-pro-495-pcs.html) — 2026-10-15
- [NVIDIA – Jetson Thor specs (128GB, 273 GB/s, 40–130W)](https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/) — 2026-10-15
- [Micro Center – Jetson AGX Thor Developer Kit ($5,499)](https://www.microcenter.com/product/711571/nvidia-jetson-agx-thor-developer-kit) — 2026-10-15
- [CNX Software – NVIDIA raises Jetson prices up to 101% (2026-07-22)](https://www.cnx-software.com/2026/07/22/nvidia-increases-the-price-of-jetson-modules-and-devkits-by-up-to-101/) — 2026-07-22
- [ComputingForGeeks – Jetson Orin Nano Super vs AI HAT+ 2 vs Hailo-10H (prices/benchmarks checked 2026-10-12)](https://computingforgeeks.com/edge-llm-jetson-orin-vs-ai-hat-2-vs-hailo-10h/) — 2026-10-12
- [NVIDIA – Jetson Orin Nano Super Developer Kit](https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/nano-super-developer-kit/) — 2026-10-15
- [oMLX benchmark – gpt-oss-120b MXFP4-Q8 on M5 Max (40c) 128GB](https://omlx.ai/benchmarks/dgadv0k9) — 2026-10-15
- [oMLX benchmark – gpt-oss-120b on M5 Ultra (80c) 256GB](https://omlx.ai/benchmarks/performance/bcoqxlcv) — 2026-10-15
- [oMLX benchmark – gpt-oss-20b MXFP4-Q8 on M5 Pro (20c) 48GB](https://omlx.ai/benchmarks/performance/h03e0nvx) — 2026-10-15
- [oMLX benchmark – gpt-oss-20b on M5 (10c) 32GB](https://omlx.ai/benchmarks/performance/xns0rgnb) — 2026-10-15
- [llamaperf – gpt-oss community reports (incl. bestllmfor M5 Max 128GB Ollama run 2026-09-16; nishtahir EVO-X2 llama-bench)](https://llamaperf.com/model/gpt-oss) — 2026-10-15
- [llm-speed – gpt-oss-20b signed runs (RTX 5090, RTX 4090, M3 Ultra)](https://llm-speed.com/m/gpt-oss-20b) — 2026-10-15
- [Hardware Corner – gpt-oss-20b on RTX 3090 (Linux, llama.cpp, FA)](https://www.hardware-corner.net/guides/gpt-oss-20b-gpu-benchamrks/) — 2025-10-26

What to run it on.

gpt-oss-120b6 published runs · decode tok/s, higher is faster

gpt-oss-20b10 published runs · decode tok/s, higher is faster

`gpt-oss-120b`6 published runs · decode tok/s, higher is faster

`gpt-oss-20b`10 published runs · decode tok/s, higher is faster