Open-Weight vs Closed Models: VRAM and RAM Requirements, Measured

The VRAM math of open-weight vs closed is really one question: "who pays for the memory?", and the decisive fact is that closed models (GPT-5, Claude, Gemini) have zero local VRAM — their weights are not published, so you cannot self-host them at all; you rent them via API and the provider owns the hardware (per-token billing, data leaves); conversely open means you buy the VRAM, and a 7B is about 5GB at Q4 but a frontier DeepSeek 671B is about 400GB even at 4-bit (8xH100), so the VRAM question turns into which open model you can host and whether you should buy or rent (an honest premise: a self-hosted open model does not replace GPT-5 or Claude — agentic and tool-use gaps remain).

In plain terms: closed is a taxi, open is buying a car. A taxi means no car (VRAM) purchase, just fares (tokens) - but you only ride the finest car (frontier) as a taxi. Buy a car (open) and you cover the parking (VRAM), fuel, and maintenance (electricity, ops), and it pays off only if you drive a lot.

Why is closed's local VRAM zero?#

Because without weights you cannot run it yourself. GPT-5, Claude, and Gemini have no open weights, so they are API-only - no local hardware, and VRAM, electricity, and ops are all the provider's. In exchange come per-token cost and data leaving your walls (Claude Opus measures $15 input / $75 output per million tokens, GPT-5.2 $1.75 / $14). The diagram below is the "taxi (rent) vs buy-a-car (self-host)" cost structure.

That is, closed is "rent capability by the hour instead of buying memory." So the self-host and VRAM discussion only holds on the open side from the start.

How much does open cost?#

The model tier is the VRAM tier. Measured: 7B Q4 about 5GB (8GB), 70B Q4 about 42GB (48GB unified or dual 3090), 405B Q4 about 243GB (non-consumer), and DeepSeek 671B - even as MoE it must load fully - about 400GB at 4-bit AWQ (8xH100, 640GB). The diagram below is that VRAM ladder - each tier up leaves consumer GPUs behind.

At FP16, 671B is a staggering 1,342GB (1.4TB). Tip: Llama 3.3 70B matches 405B on most benchmarks, standing in for the 405B tier at 43GB. MoE offloads to CPU to fit 24GB but throughput collapses. So matching a frontier open model locally needs datacenter GPUs.

Buy or rent?#

Volume and the model you compare against set the break-even. Roughly, 0.5M tokens/day crosses break-even versus frontier APIs, but versus budget APIs (open hosting on Groq/Together) you need 100M tokens/day to win (providers run hardware near full utilization). Below is that break-even crossover.

And hidden costs are large - raw GPU is only 30-40% of true cost (a 2.5-3x multiplier), and engineering labor dominates (a "free" open model runs $500K+/year). So for most teams, APIs are cheaper on a full-cost basis. Self-hosting wins when replacing an expensive flagship at high volume, or when privacy/compliance forbids the API.

So what do you pick?#

The key is hybrid, placed by volume and privacy.

• Default: closed API for exploration, low volume, and frontier needs (zero local VRAM, fast start).
• Self-host: for high volume, privacy, or fine-tuning, go open - set the tier by VRAM budget (7B 5GB to 70B 42GB) and compute your own break-even.
• Mix: self-hosted open for predictable workloads, API for flexibility and frontier. Measure exact cost on your own volume and scenarios.

Related reading: 오픈웨이트 vs 클로즈드, 5분 시작 가이드(초보자용), 에이전트 브라우저 제어 — 우리는 이렇게 운영한다(회고)

Related reading: 로컬 음성합성(TTS) 오픈모델 — VRAM·RAM 요구량 실측, Ollama·LM Studio·llama.cpp 실행기 VRAM·RAM 실측

Reference links

• Llama models (open weights)
• DeepSeek-V3 (671B MoE, open)
• vLLM (high-performance serving, quantization)
• llama.cpp (GGUF, offload)
• Qwen3 (self-host-friendly)

Note: GB, token, and price figures are public 2026 measurements and vendor data and vary by model, quantization, GPU, and API price (not permanent numbers). Reported break-evens range widely (0.5M-190M tokens/day) by GPU config, comparison price, and whether labor is included, so model your own three scenarios (these numbers are only a start). Open does not replace frontier closed. Models and prices move fast, so this is reviewed quarterly.