⌘K
170 products · 144 models · 2026 telemetry suite

Find the optimal hardware for the AI models you actually run.

The machine-readable hardware intelligence platform. Sourced benchmark telemetry, VRAM memory requirements, quantization mathematics, and performance-per-dollar rankings for GPUs, mini PCs, and workstations.

Quick jumps RTX 5090 Blackwell Llama 70B fit Used RTX 3090 Flux.1 [dev] 24 GB+ GPUs
Hardware indexed
170products

NVIDIA · AMD · Intel · Apple

AI models analyzed
144architectures

Dense · MoE · Diffusion · Audio

Sourced benchmarks
100measurements

100% provenance cited

Deep guides
25published

Answer-first methodology

Model intelligence

VRAM requirements & GPU fit, per model

Every model page states its memory math at each quantization level and links the hardware that actually runs it — server-rendered, machine-readable, and source-pinned.

LLM 2,779.9B params

Kimi K3

Kimi K3 is a mixture-of-experts model released with open weights on Hugging Face: 2779.93B total parameters, a 1,048,576-token context window, other l...

Min: 2,048 GB+ memory (Q4-class, calculated) Rec: 2,048 GB+ for comfortable context headroom (Q4-class, calculated)
Model requirements & fit →
LLM 2,446.2B params

Qwen3.8 2.4T A95B

Qwen3.8 2.4T A95B is a mixture-of-experts model with 95B activated parameters per token and 2.4T-class total open weights (official card: Number of Pa...

Min: 64 GB+ memory (Q4-class on 95B active, calculated) Rec: 96 GB+ for comfortable context headroom (Q4-class on 95B active, calculated)
Model requirements & fit →
LLM 1,600.8B params

Nex N2.5 Max

Nex N2.5 Max is a mixture-of-experts model activating 6 of 384 routed experts per token (plus 1 shared) released with open weights on Hugging Face: 16...

Min: 1024 GB+ memory (Q4-class, calculated) Rec: 1536 GB+ for comfortable context headroom (Q4-class, calculated)
Model requirements & fit →
LLM 1,600B params

DeepSeek V4 Pro

DeepSeek V4-Pro is a mixture-of-experts model with 1.6T total parameters and 49B activated per token (official README table; MIT). Config (deepseek-ai...

Min: Multi-node / ~1056 GB at Q4_K_M (calculated from 1.6T resident; T08) Rec: Multi-node data-center; not a single unified-memory Mac/Strix fit
Model requirements & fit →
IMAGE 12B params

FLUX.1 dev

FLUX.1 dev needs about 10 GB of VRAM at Q4/NF4 quantization for 1024×1024 image generation, so a 12 GB GPU such as the GeForce RTX 3060 12GB or RTX 40...

Min: 12 GB VRAM GPU at NF4/GGUF Q4_K_M (e.g. RTX 3060 12GB) Rec: 16 GB VRAM GPU (RTX 4060 Ti 16GB, RTX 4080) at FP8 for fast, high-quality generation
Model requirements & fit →
AUDIO 1.5B params

Whisper large-v3

Whisper large-v3 runs locally on almost any modern GPU: the 1.54-billion-parameter model needs about 4.4 GB of VRAM at FP16 and about 2.2 GB at Q5, in...

Min: Any 4 GB VRAM GPU at FP16 (or CPU-only via whisper.cpp) Rec: 8 GB VRAM GPU (RTX 3060/4060 class) for faster-than-real-time batch transcription
Model requirements & fit →
Answer-first guides

Explore by AI workload

Direct answers for humans and AI search systems. Every guide states top pick, best value, and budget options before the in-depth technical analysis.

Llama · DeepSeek · Qwen

Local LLM Inference

Run 7B to 70B+ models locally. Tokens/sec speed, context limits, and zero-offload VRAM tiers.

Read guide → Flux.1 · SDXL · ComfyUI

Diffusion & Image Gen

Flux.1 [dev] and SDXL throughput, batch scalability, and FP16 vs NF4 requirements.

Read guide → Whisper v3 · CoreML

Speech & Audio AI

Whisper Large v3 batch transcription. Low-power, always-on edge setups under $500.

Read guide → Interactive calculator

VRAM Memory Matrix

Mathematical calculator for weight memory, KV-cache scaling, and activation buffers.

Open calculator →
Telemetry leaderboard

Hardware rankings & benchmark index

Computed live from database records. Spec ratios are labeled as spec proxies with their formulas — measured throughput appears only where sourced benchmark rows exist.

All leaderboards →

Memory bandwidth per $1k

GB/s · $ →
  1. #1 2,860
  2. #2 1,831
  3. #3 1,772
  4. #4 1,735
  5. #5 1,605

VRAM per $100

GB · $ →
  1. #1 9.3
  2. #2 5.1
  3. #3 4.9
  4. #4 4.8
  5. #5 4.6

Llama 3.1 8B @ Q4_K_M

tok/s
  1. #1 215
  2. #2 132
  3. #3 125
  4. #4 110
  5. #5 102

* = community estimate, not measured

Measured Llama 3.1 8B Q4_K_M token generation (tok/s) — sourced benchmark rows only; estimates marked.
Rank Hardware VRAM Bandwidth MSRP Throughput Action
#1 GeForce RTX 5090 NVIDIA 32 GB 1,792 GB/s $1,999 215 tok/s Details
#2 GeForce RTX 5080 NVIDIA 16 GB 960 GB/s $999 132 tok/s Details
#3 GeForce RTX 4090 NVIDIA 24 GB 1,008 GB/s $1,599 125 tok/s Details
#4 RTX 6000 Ada NVIDIA 48 GB 960 GB/s $6,800 110 tok/s Details
#5 GeForce RTX 4080 SUPER NVIDIA 16 GB 736 GB/s $999 102 tok/s Details
#6 GeForce RTX 4070 SUPER NVIDIA 12 GB 504 GB/s $599 92 tok/s Details
#7 GeForce RTX 3090 NVIDIA 24 GB 936 GB/s $1,499 85 tok/s Details
#8 Radeon RX 7900 XTX AMD 24 GB 960 GB/s $999 76 tok/s Details
Fresh telemetry

Latest indexed hardware

Full database →
Apple 2026

Apple M6

memory TBD
Apple 2026

Apple M5 Ultra

512 GB memory
Apple 2026

Apple M5 Pro

64 GB memory
Apple 2026

Apple M5 Max

128 GB memory
Editorial intelligence

Buying guides

All guides →
Editorial integrity

Every number has a source. Every estimate is labeled.

We distinguish tested lab benchmarks from community measurements and mathematical estimations. Our formulas are open, documented, and fully machine-readable for AI search crawlers.

Read our verification & testing methodology →
Independent benchmark

Replicated press/lab numbers with cited methodology.

Community verified

Multiple consistent owner reports; never presented as measured.

Calculated estimate

Open-formula math from bandwidth and memory specs.

Vendor spec

Manufacturer datasheets — labeled as vendor numbers.