Inference Memory & Cost Modeling
Local LLM VRAM & inference speed calculator
Calculate whether open-weights models fit your GPU VRAM or Apple Mac unified memory, estimate generation speeds (tokens/sec) from memory bandwidth, and compute break-even points versus frontier API costs.
! PARTIAL OFFLOAD25.22 GB (105%)~7 t/s
Qwen 2.5 Coder 32B Instruct on NVIDIA GeForce RTX 4090 (24GB GDDR6X)
! Partial Offload — Slow CPU/RAM Spillover
! PARTIAL OFFLOAD
Required VRAM
25.22 GB
105% of hardware capacity (24 GB)Est. Decode Speed
~7 tok/s
Dense model @ 1008 GB/sAvailable Headroom
-1.22 GB
Needs +1.22 GB more VRAMMemory Allocation & Headroom
25.22 GB allocated of 24 GB capacity
105% utilized
Weights: 19.5 GB
KV: (32K): 4.29 GB
Overhead: 1.43 GB
Deficit: -1.22 GB
Architecture & Throughput Transparency: Single-user decode speed (~7 tok/s) is memory-bandwidth bound at batch size = 1 on NVIDIA GeForce RTX 4090 (24GB GDDR6X) (1008 GB/s). Prompt prefill / TTFT (~102 tok/s) is compute-bound and utilizes GPU Tensor cores. Actual latency varies with FlashAttention-2 and runtime engine (Ollama/vLLM/llama.cpp).
Local Hardware Configurations & Pricing
Partial Offload: Dedicated 100% GPU OptionsConsumer GPU✓ Fits Model
32 GB VRAMNVIDIA GeForce RTX 5090 (32GB GDDR7)
Maximum tokens/sec on 27B–35B models & Blackwell FP4 quantization
1792 GB/s575W TDP
$1,999Amazon Prime →
Consumer GPU
24 GB VRAMNVIDIA GeForce RTX 4090 (24GB GDDR6X)
The 24GB CUDA enthusiast standard for 14B–32B Q4_K_M models
1008 GB/s450W TDP
$1,699Amazon Prime →
Multi-GPU Rig✓ Fits Model
48 GB VRAM2x NVIDIA RTX 3090 Dual Rig (48GB Total VRAM)
Best price-to-VRAM homelab rig for Llama 3.3 70B Q4_K_M & Qwen 2.5 Coder 32B Q8_0
1648 GB/s700W TDP
$1,550Amazon Prime →
Consumer GPU
16 GB VRAMNVIDIA GeForce RTX 5080 (16GB GDDR7)
Blazing fast 8B–14B coding models with FP4/Q6_K precision
960 GB/s360W TDP
$999Amazon Prime →
Out of VRAM? Don't downgrade model quality.
Out of VRAMSpin up an on-demand cloud GPU instance in 60 seconds with pre-installed Ollama and vLLM. (Qwen 2.5 Coder 32B Instruct @ 25.22 GB VRAM)
RunPod Cloud GPUsBest 1-Click Ollama & vLLM
$0.38/hr (RTX 4090 Spot) – $2.39/hr (H100)Vast.ai GPU MarketplaceCheapest Dual RTX 3090 / 4090
$0.18/hr (RTX 3060) – $0.78/hr (2x RTX 3090)Lambda GPU CloudEnterprise 8x H100 Clusters
$2.49/hr (H100) – $19.92/hr (8x H100)Quantization Precision Matrix
32K ctx • KV: Q8_0| Quantization | Weights | KV Cache | Total VRAM | Est. Speed | Fit Status |
|---|---|---|---|---|---|
| FP1616 bpw | 65 GB | 4.29 GB | 72.32 GB (301%) | 0 t/s | OOM |
| Q8_08.5 bpw | 34.45 GB | 4.29 GB | 40.7 GB (170%) | 0 t/s | OOM |
| Q6_K6.56 bpw | 26.65 GB | 4.29 GB | 32.62 GB (136%) | 0 t/s | OOM |
| Q5_K_M5.68 bpw | 23.07 GB | 4.29 GB | 28.92 GB (121%) | ~6 t/s | PARTIAL |
| Q4_K_M4.8 bpw | 19.5 GB | 4.29 GB | 25.22 GB (105%) | ~7 t/s | PARTIAL |
| FP44.25 bpw | 17.23 GB | 4.29 GB | 22.87 GB (95%) | ~38 t/s | FITS |
| IQ3_XS3.3 bpw | 13.32 GB | 4.29 GB | 18.83 GB (78%) | ~48 t/s | FITS |
Export Full Architecture & Infrastructure Spec
Download an instant Markdown architecture spec with exact memory math, hardware benchmarks, and runtime CLI flags.
Pre-Calculated Benchmark Directory
370+ Pre-computed hardware compatibility matrices and speed benchmarksBest Local LLMs for 12GB VRAM GPUs12GB VRAMBest Local LLMs for 16GB VRAM GPUs16GB VRAMBest Local LLMs for 24GB VRAM GPUs24GB VRAMBest Local LLMs for 32GB VRAM32GB VRAMBest Local LLMs for 48GB VRAM48GB VRAMBest Local LLMs for 64GB Unified Memory64GB VRAMBest Local LLMs for 128GB Unified Memory128GB VRAM