← Back to Will-it-run

Custom build engine

Describe your build — any GPUs, CPU, RAM, OS, runtime, use case. We'll compute effective VRAM honestly, recommend a runtime, and tell you which models fit comfortably, which are borderline, and which aren't practical.

Total VRAM ≠ pooled VRAM. We never sum VRAM unless the silicon truly pools (Apple unified memory). We always explain why effective is lower than total.

Calculations follow the RunLocalAI Will-It-Run Framework: effective VRAM, model working set, runtime constraints, fit tiers, and measured-vs-estimated evidence labels.

Describe your build

Add GPUs, set CPU/RAM/OS, optionally pick a runtime + use case. URL updates as you change fields — share a build by copying the URL.

Build summary

Total VRAM
192 GB
Effective VRAM
~150 GB
range 142-154 GB
Topology
single node multi gpu
pcie
Setup difficulty
advanced
speed penalty ~18%
Why effective VRAM is lower than total

8× NVIDIA GeForce RTX 4090 = 192 GB total VRAM, but without NVLink, cross-card bandwidth is PCIe-bound (~32 GB/s vs NVLink ~112 GB/s). With tensor-parallelism, each card holds ~1/8 of the model weights and replicates activations + KV cache. After 15% TP overhead, effective model capacity is ~150 GB. Largest single tensor on one card is ~22 GB.

Measured evidence on this hardware

Publicly inspectable measured rows for the selected hardware slug(s). Exact measured rows calibrate the fit table instead of leaving it as pure VRAM estimation.

No publicly inspectable benchmark rows are attached to this exact hardware yet. The engine will still calculate fit and runtime, but speed rows will remain estimated.

Recommended runtime

Best engine for this topology + skill level + use case.

Ray Serve
primary
involved

Ray Serve is the runtime you selected. It is compatible enough for this build, so the custom engine keeps it as the primary path while still showing alternatives.

vLLM
primary
involved

Tensor-parallel across NVLink/PCIe — works on every recent consumer + datacenter pair. AWQ-INT4 + 70B fits dual 3090 / dual 4090 cleanly.

ExLlamaV2
alternative
involved

Single-stream king. EXL2 4.0bpw + 70B fits dual 3090 with NVLink and beats vLLM on solo-user throughput.

WORKLOAD PROFILE
FITS
Command R+ (Aug 2024) @ AWQ-INT4, 8K context on NVIDIA GeForce RTX 4090
0 GB150 GBVRAM ceiling
Weights52 GB
KV cache52 GB
Activations2.6 GB
Runtime1.8 GB
Headroom42 GB
ESTIMATED DECODE RATE
11 tok/s
Bandwidth-derived estimate · efficiency 0.55. Real-world rates land within ±20% on well-tuned runtimes.
11 tokens per second02550100150

Models that fit your build

177 models considered. Categorized by headroom at the recommended quant + a sensible context for your use case.

Comfortable
24 models · ≥15% headroom
ModelParamsQuantVRAM est.ContextEvidenceNote
Command R+ (Aug 2024)104BAWQ-INT4106.6 GB8,192No measured row yetFits cleanly at AWQ-INT4 + 8,192 ctx with 29% headroom.
Command R+ 104B104BQ4_K_M115 GB8,192No measured row yetFits cleanly at Q4_K_M + 8,192 ctx with 23% headroom.
Llama 3.2 90B Vision90BAWQ-INT492.3 GB8,192No measured row yetFits cleanly at AWQ-INT4 + 8,192 ctx with 38% headroom.
Llama 3.2 90B Vision Instruct90BQ4_K_M98.6 GB8,192No measured row yetFits cleanly at Q4_K_M + 8,192 ctx with 34% headroom.
InternVL 2.5 78B78BQ4_K_M86.3 GB8,192No measured row yetComfortable fit with 42% headroom — room to extend context or run alongside other workloads.
Qwen 2.5 72B Instruct72BQ5_K_M87.5 GB8,192No measured row yetComfortable fit with 42% headroom — room to extend context or run alongside other workloads.
Qwen 2.5 Math 72B72BQ4_K_M61.1 GB4,096No measured row yetComfortable fit with 59% headroom — room to extend context or run alongside other workloads.
Qwen 2.5-VL 72B72BAWQ-INT480.1 GB8,192No measured row yetComfortable fit with 47% headroom — room to extend context or run alongside other workloads.
DeepSeek R1 Distill Llama 70B70BQ5_K_M84.4 GB8,192No measured row yetComfortable fit with 44% headroom — room to extend context or run alongside other workloads.
Llama 3.1 70B Instruct70BQ5_K_M84.4 GB8,192No measured row yetComfortable fit with 44% headroom — room to extend context or run alongside other workloads.
Llama 3.3 70B Instruct70BQ8_076.2 GB8,192No measured row yetComfortable fit with 49% headroom — room to extend context or run alongside other workloads.
Hermes 3 Llama 3.1 70B70BQ4_K_M77 GB8,192No measured row yetComfortable fit with 49% headroom — room to extend context or run alongside other workloads.
OpenBioLLM Llama 3 70B70BQ4_K_M79.1 GB8,192No measured row yetComfortable fit with 47% headroom — room to extend context or run alongside other workloads.
Tulu 3 70B70BQ4_K_M80.2 GB8,192No measured row yetComfortable fit with 47% headroom — room to extend context or run alongside other workloads.
Llama 3.1 Nemotron 70B Instruct70BQ4_K_M79.6 GB8,192No measured row yetComfortable fit with 47% headroom — room to extend context or run alongside other workloads.
Jamba 1.5 Mini52BQ4_K_M58.6 GB8,192No measured row yetComfortable fit with 61% headroom — room to extend context or run alongside other workloads.
Mixtral 8x7B Instruct47BQ5_K_M58.2 GB8,192No measured row yetComfortable fit with 61% headroom — room to extend context or run alongside other workloads.
Command R 35B35BQ4_K_M39.6 GB8,192No measured row yetComfortable fit with 74% headroom — room to extend context or run alongside other workloads.
Aya 23 35B35BQ4_K_M39.6 GB8,192No measured row yetComfortable fit with 74% headroom — room to extend context or run alongside other workloads.
Yi 1.5 34B34BQ4_K_M29.5 GB4,096No measured row yetComfortable fit with 80% headroom — room to extend context or run alongside other workloads.
Phind CodeLlama 34B v234BQ4_K_M38 GB8,192No measured row yetComfortable fit with 75% headroom — room to extend context or run alongside other workloads.
Aya Expanse 32B32BAWQ-INT432.8 GB8,192No measured row yetComfortable fit with 78% headroom — room to extend context or run alongside other workloads.
Qwen 2.5 32B Instruct32BQ8_051.7 GB8,192No measured row yetComfortable fit with 66% headroom — room to extend context or run alongside other workloads.
Qwen 2.5 Coder 32B Instruct32BQ8_037.9 GB8,192No measured row yetComfortable fit with 75% headroom — room to extend context or run alongside other workloads.
Borderline
2 models · tight, may need quant downgrade
ModelParamsQuantVRAM est.ContextEvidenceNote
Mistral Large 2 (123B)123BQ4_K_M138.2 GB8,192No measured row yetTight fit at Q4_K_M — only 8% headroom. KV cache for longer context will OOM. Cap context tighter or drop one quant level.
Llama 4 Scout109BQ5_K_M136.4 GB8,192No measured row yetTight fit at Q5_K_M — only 9% headroom. KV cache for longer context will OOM. Cap context tighter or drop one quant level.
Not practical
15 models · oversize for this build
ModelParamsQuantVRAM est.ContextEvidenceNote
Molmo 72B72BF32325.7 GB4,096No measured row yet~325.7 GB needed at F32 + 4,096 ctx — overshoots effective VRAM by 117%. Drop quant or move to a larger build.
Mixtral 8x22B Instruct141BQ4_K_M158.7 GB8,192No measured row yet~158.7 GB needed at Q4_K_M + 8,192 ctx — overshoots effective VRAM by 6%. Drop quant or move to a larger build.
Qwen 3 235B-A22B235BQ4_K_M266.6 GB8,192No measured row yet~266.6 GB needed at Q4_K_M + 8,192 ctx — overshoots effective VRAM by 78%. Drop quant or move to a larger build.
DeepSeek V2.5 236B236BQ4_K_M258.7 GB8,192No measured row yet~258.7 GB needed at Q4_K_M + 8,192 ctx — overshoots effective VRAM by 72%. Drop quant or move to a larger build.
DeepSeek Coder V2 236B236BQ4_K_M258.7 GB8,192No measured row yet~258.7 GB needed at Q4_K_M + 8,192 ctx — overshoots effective VRAM by 72%. Drop quant or move to a larger build.
Llama 3.1 Nemotron Ultra 253B253BQ4_K_M277.7 GB8,192No measured row yet~277.7 GB needed at Q4_K_M + 8,192 ctx — overshoots effective VRAM by 85%. Drop quant or move to a larger build.
DeepSeek V4 Flash (284B MoE)284BQ4_K_M312.1 GB8,192No measured row yet~312.1 GB needed at Q4_K_M + 8,192 ctx — overshoots effective VRAM by 108%. Drop quant or move to a larger build.
Hunyuan Large 389B MoE389BQ3_K_M393.6 GB8,192No measured row yet~393.6 GB needed at Q3_K_M + 8,192 ctx — overshoots effective VRAM by 162%. Drop quant or move to a larger build.
Jamba 1.5 Large398BQ4_K_M451 GB8,192No measured row yet~451.0 GB needed at Q4_K_M + 8,192 ctx — overshoots effective VRAM by 201%. Drop quant or move to a larger build.
Llama 4 Maverick400BQ4_K_M452 GB8,192No measured row yet~452.0 GB needed at Q4_K_M + 8,192 ctx — overshoots effective VRAM by 201%. Drop quant or move to a larger build.
DeepSeek R1 (671B reasoning)671BQ4_K_M734.5 GB8,192No measured row yet~734.5 GB needed at Q4_K_M + 8,192 ctx — overshoots effective VRAM by 390%. Drop quant or move to a larger build.
DeepSeek V3 (671B MoE)671BQ4_K_M734.5 GB8,192No measured row yet~734.5 GB needed at Q4_K_M + 8,192 ctx — overshoots effective VRAM by 390%. Drop quant or move to a larger build.
Ring-2.6-1T1000BQ3_K_M1011.9 GB8,192No measured row yet~1011.9 GB needed at Q3_K_M + 8,192 ctx — overshoots effective VRAM by 575%. Drop quant or move to a larger build.
Kimi K2.61000BQ4_K_M1130 GB8,192No measured row yet~1130.0 GB needed at Q4_K_M + 8,192 ctx — overshoots effective VRAM by 653%. Drop quant or move to a larger build.
DeepSeek V4 Pro (1.6T MoE)1600BQ4_K_M1766 GB8,192No measured row yet~1766.0 GB needed at Q4_K_M + 8,192 ctx — overshoots effective VRAM by 1077%. Drop quant or move to a larger build.
Rent instead
Not enough VRAM for this model? Rent a GPU by the hour

This build can't hold Molmo 72B. By our estimate, a rented machine with at least 326 GB of GPU memory in total (at this size, usually several GPUs) runs it now, without buying new hardware.

Referral links: RunLocalAI may earn a commission or account credits if you sign up and rent, at no extra cost to you. Hourly rates and which GPUs are listed change all the time, so check the provider's page before you start. How we make money.

Related

Multi-GPU buying guide →

NVLink vs PCIe, tensor- vs pipeline-parallel, mixed-card honesty.

Hardware combinations →

Curated multi-GPU / cluster setups with effective-VRAM math.

Setup path-finder →

OS + runtime install commands for your stack.

Compatibility matrix →

Runtime × OS × hardware support truth table.

Shopping a full build instead of a single card?

If you're sizing a fresh AI build (not just a card to drop into an existing system), the build-budget walkthroughs cover the whole BOM honestly: AI PC build under $1,000 or AI PC build under $2,000 cover the realistic 2026 budget tiers.

Vertical-fit shopping? AI PC for students covers the budget + portability tradeoffs; AI PC for developers covers the coding workflow specifics; AI PC for small business covers the document-RAG / always-on machine.

Form-factor first? See best laptop for local AI, best Mac for local AI, best mini PC for local AI, or best used GPU for local AI.