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.
How two used RTX 3090 cards hold a 70B Q4_K_M coding/reasoning model completely inside high-speed GDDR6X.
OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=q8_0 \ ollama run llama3.3:70b-instruct-q4_k_m --ctx-size 8192
| Quantization | Memory Needed | Speed Est. | Placement |
|---|---|---|---|
| FP16 / BF16 (16-bit)Reference Lossless | 143.65 GB | OOM | OOM |
| Q8_0 (8-bit)Near-Lossless | 77.46 GB | 1.7 tok/s | RAM Offload |
| Q6_K (6-bit K-Quant)Studio Fidelity | 60.34 GB | 2.6 tok/s | RAM Offload |
| Q5_K_M (5-bit Medium)Minimal Loss | 52.75 GB | 3.4 tok/s | RAM Offload |
| Q4_K_M (4-bit Sweet Spot)Recommended Default | 45.25 GB | 4.5 tok/s | RAM Offload |
| IQ3_XS (3.3-bit I-Quant)Noticeable Degradation | 32.46 GB | 39.3 tok/s | Pure GPU VRAM |
| IQ2_XXS (2.1-bit Extreme)Emergency Fit Only | 21.87 GB | 60.6 tok/s | Pure GPU VRAM |
| Model | Best Pure-VRAM Quant | Memory | Est. Speed |
|---|---|---|---|
| Llama 3.1 8B InstructUltra-fast RAG extraction, IDE autocomplete & classification | FP16 | 17.46 GB | 73.6 tok/s |
| Qwen 2.5 Coder 14BMid-tier local VS Code copilot balancing speed & repo accuracy | FP16 | 31.05 GB | 40.3 tok/s |
| Mistral Small 24BClean instruction following on 24 GB single-GPU workstations | Q8_0 | 27.03 GB | 46.4 tok/s |
| Gemma 2 / 3 27BMultilingual & editorial writing in single 24–32 GB GPUs | Q8_0 | 31.24 GB | 40.8 tok/s |
| Qwen 2.5 32B / DeepSeek-R1 32BGold-standard local coding & reasoning on RTX 3090/4090/5090 & 48GB Macs | Q8_0 | 36.43 GB | 34.3 tok/s |
| Mixtral 8x7B MoE (46.7B Total)Sparse Mixture-of-Experts: needs 47B RAM footprint, computes at 13B speed | Q6_K | 39.99 GB | 111.5 tok/s |
| Llama 3.3 70B / DeepSeek-R1 70BGPT-4 class private local reasoning for 48 GB–128 GB rigs | IQ3_XS | 32.46 GB | 39.3 tok/s |
| DeepSeek R1 671B MoE (Full Frontier)Cluster or Mac Studio Ultra / dual EPYC server deployment (37B active per token) | Needs Offload / Cluster | 408.95 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 | Q8_0 | 33.68 GB | 333.5 tok/s |
| Qwen3-32B (Dense Dual-Thinking)2026 dense local reasoning & coding benchmark with dual thinking/non-thinking toggle | Q8_0 | 36.75 GB | 33.9 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 | IQ2_XXS | 31.93 GB | 248.1 tok/s |
Pre-configured telemetry for Llama 3.3 70B / DeepSeek-R1 70B (Q4_K_M) on 2x NVIDIA RTX 3090 NVLink (48 GB). Requires 45.25 GB (42.8 GB weights + 1.25 GB Q8 KV cache) against 45.12 GB usable pool for ~4.5 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.