Cloud vs. Rig TCO
Plot cumulative 36-month API spending against hourly RunPod/H100 rentals and owned RTX/Mac hardware with electricity and resale value.
Open Workbench
Compare 8B Q8_0 vs. 14B Q4_K_M weights inside an affordable 12 GB GPU.
OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=q8_0 \ ollama run llama3.1:8b-instruct-q8_0 --ctx-size 16384
| Quantization | Memory Needed | Speed Est. | Placement |
|---|---|---|---|
| FP16 / BF16 (16-bit)Reference Lossless | 17.96 GB | 2.4 tok/s | RAM Offload |
| Q8_0 (8-bit)Near-Lossless | 10.43 GB | 26.9 tok/s | Pure GPU VRAM |
| Q6_K (6-bit K-Quant)Studio Fidelity | 8.48 GB | 38.5 tok/s | Pure GPU VRAM |
| Q5_K_M (5-bit Medium)Minimal Loss | 7.62 GB | 44.1 tok/s | Pure GPU VRAM |
| Q4_K_M (4-bit Sweet Spot)Recommended Default | 6.77 GB | 51.7 tok/s | Pure GPU VRAM |
| IQ3_XS (3.3-bit I-Quant)Noticeable Degradation | 5.31 GB | 72.8 tok/s | Pure GPU VRAM |
| IQ2_XXS (2.1-bit Extreme)Emergency Fit Only | 4.11 GB | 109.9 tok/s | Pure GPU VRAM |
| Model | Best Pure-VRAM Quant | Memory | Est. Speed |
|---|---|---|---|
| Llama 3.1 8B InstructUltra-fast RAG extraction, IDE autocomplete & classification | Q8_0 | 10.43 GB | 26.9 tok/s |
| Qwen 2.5 Coder 14BMid-tier local VS Code copilot balancing speed & repo accuracy | IQ3_XS | 8.65 GB | 40 tok/s |
| Mistral Small 24BClean instruction following on 24 GB single-GPU workstations | IQ2_XXS | 8.75 GB | 38.2 tok/s |
| Gemma 2 / 3 27BMultilingual & editorial writing in single 24–32 GB GPUs | IQ2_XXS | 11.26 GB | 29.5 tok/s |
| Qwen 2.5 32B / DeepSeek-R1 32BGold-standard local coding & reasoning on RTX 3090/4090/5090 & 48GB Macs | Needs Offload / Cluster | 22.6 GB | 1.9 tok/s |
| Mixtral 8x7B MoE (46.7B Total)Sparse Mixture-of-Experts: needs 47B RAM footprint, computes at 13B speed | Needs Offload / Cluster | 30.51 GB | 4.7 tok/s |
| Llama 3.3 70B / DeepSeek-R1 70BGPT-4 class private local reasoning for 48 GB–128 GB rigs | Needs Offload / Cluster | 46.5 GB | 0.8 tok/s |
| DeepSeek R1 671B MoE (Full Frontier)Cluster or Mac Studio Ultra / dual EPYC server deployment (37B active per token) | Needs Offload / Cluster | 409.9 GB | OOM |
| Qwen3-30B-A3B (MoE — 3.3B Active)Ultra-fast 30B Sparse MoE decoding at 3.3B speed; effortless fit on 24 GB cards | IQ2_XXS | 10.04 GB | 254.1 tok/s |
| Qwen3-32B (Dense Dual-Thinking)2026 dense local reasoning & coding benchmark with dual thinking/non-thinking toggle | Needs Offload / Cluster | 22.78 GB | 1.9 tok/s |
| Llama 4 Scout 109B (MoE — 17B Active)Multimodal 109B MoE (17B active) built for 64–128 GB Macs, Dual-3090, or H100 pods | Needs Offload / Cluster | 68.78 GB | OOM |
Pre-configured telemetry for Llama 3.1 8B Instruct (Q8_0) on NVIDIA RTX 3060 (12 GB). Requires 10.43 GB (8.53 GB weights + 1 GB Q8 KV cache) against 11.28 GB usable pool for ~26.9 tok/s decode.
Primary Technical & US Grid Benchmarks
Plot cumulative 36-month API spending against hourly RunPod/H100 rentals and owned RTX/Mac hardware with electricity and resale value.
Calculate daily, monthly, and annual kWh bills using 80+ PSU efficiency, duty cycle, optional AC cooling load, and 50 US state EIA rates.
Combine dual-GPU draw, CPU peak boost, NVMe arrays, and ATX 3.1 transient spikes to hit the 50–70% PSU efficiency sweet spot.