Method & sources
VRAM formula
need(quant, ctx) = total_8k(quant) + kv_kib_per_token_fp16 × (ctx − 8192) ÷ 1048576
We trust the open dataset's per-model total@8k (weights + KV cache + runtime overhead, dataset methodology), and extrapolate only the KV-cache term linearly with context. That shape — weights + KV cache + runtime allowance — is the one every public local-LLM calculator uses, llamaperf.com among them (retrieved 2026-10-20); we do not mirror their numbers and we cannot audit their internals (their calculator is JS-rendered and publishes no written formula we could read), so this formula is ours and the inputs are the dataset’s.
Known discrepancy, stated openly: for DeepSeek-V4-Flash our dataset's fp8 weights are 271 GiB while vLLM official recipes, retrieved 2026-10-20: FP8 ≈ 200 GB, NVFP4 ≈ 170 GB, and Thunder Compute blog, 2026-08-12: native checkpoint ≈ 175 GB, 3-bit GGUF ≈ 131 GB. Causes: checkpoint vs runtime footprint, MoE expert bookkeeping, MLA/KV variants. Treat any single number ±10–15% and check our numbers against your real ollama/vLLM logs — your measurement is worth more than any table.
macOS usable-memory rule: ours, and how to check yours
We count GPU-usable memory as 2/3 of unified RAM (3/4 above 36GB). Be clear about what that is: our conservative operating rule, not an Apple-published constant. The mechanism is real — Metal reports a per-machine ceiling, recommendedMaxWorkingSetSize — but the ratio moves with machine, RAM size and OS version; public measurements we found on 2026-10-20 put it roughly between 66% and 75% of unified memory. So check yours in Terminal:
sysctl iogpu.wired_limit_mb # bytes are reported in MB; divide by 1024 for GiB
If your number is higher than ours, the model fits better than our table says; if it is lower, trust your machine and tell us — we would rather publish your reading than defend a rule of thumb.
Every price source on this site
| number | value(s) | source & dates |
|---|---|---|
| Cloud GPU $/h | 5090 $0.99 · A100 40GB $1.99 · A100 80GB $1.59 · H100 $3.29–3.49 · H200 $4.59 | modelvram.com public table (self-stated check 2026-09-29, retrieved 2026-10-20). Partial re-verify 2026-10-20: Lambda A100 40GB $1.99/h CONFIRMED on lambda.ai/pricing (fetched directly); Lambda H100 rows ambiguous on-page ($3.99 vs $4.09 configs) → not adopted; RunPod & Vast pricing pages are JS-rendered, no extractable static figures — verify rates visually before spending |
| RTX 4090 MSRP | $1,599 (2022-10-12) | Wikipedia GeForce RTX 40 series — prose: “released … on October 12, 2022, for $1,599 US”, retrieved 2026-10-20 |
| RTX 5090 MSRP | $1,999 | Wikipedia GeForce RTX 50 series — “significantly more expensive at $1999 than the RTX 4090, which had an MSRP of $1599”, retrieved 2026-10-20 |
| RTX 3090 MSRP | $1,499 (2020-09-24) | Wikipedia GeForce RTX 30 series spec-table MSRP column, retrieved 2026-10-20. Found on the 2026-10-20 re-attempt (R8): the earlier “not retrieved” was our parser missing the figure because table cells carry no “$”. We verified the column by cross-reading the 3080 Ti row, whose 1,199 matches the launch price stated in that article’s prose. Launch price only — today’s used street price stays user-entered. |
| VRAM GiB per model/quant | 46 models × up to 9 quants | modelvram.com open dataset (repo, CC BY 4.0; repo LICENSE MIT), retrieved 2026-10-20 |
| Wan 2.1 VRAM (8.19 / 69.1 / 76.7 / 29.9 GB) | official README bullets + efficiency figure | Wan-Video/Wan2.1 official README + efficiency figure assets/comp_effic.png (retrieved 2026-10-20); full matrix transcribed verbatim into our repo archive data/wan21_efficiency_official.txt |
| FLUX params 11,891,178,560 → bf16 floor 22.1 GiB | params official; floor = our arithmetic; official VRAM figure: does not exist | huggingface.co/api/models/black-forest-labs/FLUX.1-schnell → safetensors BF16 params 11,891,178,560, retrieved 2026-10-20; attempts 2026-10-20: HF dev card gated (401) / GitHub docs have no VRAM text / bfl.ai JS-only |
| Derived: rent cost of ONE 720P Wan 2.1 14B video all values: $1.48 / $1.52 / $1.57 / $1.66 / $1.78 / $1.83 / $1.93 / $2.23 |
A800/A100 class $1.48 (1 GPU) → $1.66 (8 GPUs) · H800/H100 class $1.78 (1 GPU) → $2.23 (8 GPUs). Derived by us from official timings — not a quoted price. | formula: official seconds ÷ 3600 × card count × same-class per-card rate (A800/A100 timings × A100-80GB rate; H800/H100 timings × H100-80GB rate). Class-matched only: pairing an H100 timing with an A100 rate understated cost by ~45% in an earlier draft of ours, so cross-class math is banned here. Card counts 1/2/4/8 are priced at the same per-card rate because multi-GPU instance pricing is not independently verified → treat as order-of-magnitude, not a quote. Timings: Wan-Video/Wan2.1 official README + efficiency figure assets/comp_effic.png (retrieved 2026-10-20); full matrix transcribed verbatim into our repo archive data/wan21_efficiency_official.txt; rates as the row above (modelvram.com public table (self-stated check 2026-09-29, retrieved 2026-10-20). Partial re-verify 2026-10-20: Lambda A100 40GB $1.99/h CONFIRMED on lambda.ai/pricing (fetched directly); Lambda H100 rows ambiguous on-page ($3.99 vs $4.09 configs) → not adopted; RunPod & Vast pricing pages are JS-rendered, no extractable static figures — verify rates visually before spending) |
| Budget tiers $500 / $900 / $1,000 / $1,500 | scenario labels — how much cash you have, not a price claim | n/a: no card, service or build on this site is claimed to cost these amounts; anything that would need a price carries a user input box instead |
| Used street prices, electricity $/kWh, API $/M tokens | user-entered | no reliable dated public source → we refuse defaults |
What we never do
No invented traffic/MRR/price numbers, no fake scarcity, `/go/` links carry disclosure + sponsored attribute.