What LLM fits my hardware?
The math behind local LLM memory budgets. Plug in a GPU and a context length; see what's actually left for weights after the KV cache, runtime overhead, and a safety margin.
Live site · Run the calculator · Read the math · Join the discussion
| GPU presets | curated models | quant formats | agent-readable routes | open source |
Every "can I run this LLM" tool lists models and calls it done. They show a green checkmark next to a 70B on a 24GB card and hope you don't ask why. vrambudget skips the magic. There is one formula and three taxes:
VRAM ~= params × (bits ÷ 8)
That's the floor. Everything else is overhead.
weights_budget = total_vram × (1 − safety)
− kv_cache × concurrency
− framework_overhead
Three subtractions and one multiplication. Anyone telling you it needs to be more complicated is selling you something.
The calculator is on the home page. Five sliders, four tiles, one bar, three tabs.
Across the top of the calculator: eight family tabs (RTX 50, RTX 40, RTX 30, Apple Silicon, Workstation, Datacenter, AMD, Intel). Each tab reveals every card in that family. Click a card to lock the calculator to its stock VRAM.
For a long-form view of every GPU at once, visit /gpu.
| Slider | What it does |
|---|---|
| VRAM | Total memory on the card. Auto-set from your GPU choice; override freely. |
| System RAM | Reserved for future model-split workflows. Not currently in the math. |
| Context | Token window. 8K is conservative; 32K and 128K make KV cache the dominant tax. |
| Concurrency | How many requests serve in parallel. Each one gets its own KV cache. |
| Safety headroom | Percentage of VRAM you refuse to spend. 15 percent is a sane default. |
- Total VRAM — device capacity
- KV cache — what the context-tax costs you
- Runtime overhead — CUDA / kernels / allocator slack
- Weights budget — what's actually left for the model
The orange tile is the answer. The budget bar below shows the four pieces stacked so you can see where the VRAM went.
Three tabs below the bar:
- Curated picks — the 30 catalog models, grouped by fit (fits / tight / over). Each row picks the highest-bit quantization that fits your budget. Click any model name to open its Hugging Face repo.
- Search Hugging Face — live search across HF's GGUF-tagged models. Type a name; results render with the same fit classification keyed to your current budget.
- By size — drag a params slider; see all nine quants in a 3×3 matrix with the size and fit class for each.
Every GPU card in the grid has a ↗ link to its detail page. Each detail page pre-binds the calculator to that card, plus a fit table of the top 12 models, plus four "compare to" cards for stepping up or down by VRAM.
Same for models: every model name links to /model/<slug> with the GPU-recommendation matrix across five quants.
RTX 50: 5070 · 5070 Ti · 5080 · 5090
RTX 40: 4060 · 4060 Ti 16GB · 4070 · 4070 Super · 4070 Ti Super · 4080 · 4080 Super · 4090
RTX 30: 3060 · 3070 · 3080 · 3080 Ti · 3090 · 3090 Ti
Apple Silicon: M2 Max 64 · M2 Ultra 192 · M3 Max 64 · M3 Max 96 · M3 Ultra 512 · M4 Pro 64 · M4 Max 128 · M5 Pro 64 · M5 Max 128
Workstation: A6000 · RTX 6000 Ada · RTX Pro 6000 · L40S
Datacenter: H100 80GB · H200 · B200 · DGX Spark · 2× H100 NVL
AMD: RX 7900 XTX · W7900 · MI300X
Intel: Arc B580 · Arc Pro B60 · Gaudi 3
Meta: Llama 3.2 1B · Llama 3.2 3B · Llama 3.1 8B · Llama 3.3 70B · Llama 3.1 405B
Alibaba: Qwen 2.5 7B · Qwen 2.5 32B · Qwen 2.5 Coder 32B · Qwen 2.5 72B · Qwen3 30B A3B · Qwen 3.5 9B · Qwen 3.6 27B · Qwen 3.6 35B A3B
OpenAI: gpt-oss 20B · gpt-oss 120B
Mistral: Mistral 7B v0.3 · Mistral Small 3 · Mixtral 8x7B · Mixtral 8x22B
Microsoft: Phi-4 Mini · Phi-4
Google: Gemma 2 9B · Gemma 4 E4B · Gemma 4 26B A4B
DeepSeek: DeepSeek V3 · DeepSeek R1
Cohere: Command R+
IBM: Granite 8B Code
BigCode: StarCoder2 15B
01.AI: Yi 34B
FP16/BF16 · FP8/INT8 · Q8_0 · Q6_K · Q5_K_M · Q4_K_M · Q3_K_M · AWQ · GPTQ
Cross-platform: Ollama · LM Studio
Apple Silicon: MLX · oMLX
Server: vLLM
LLMs ship faster than any curated list can keep up. vrambudget runs three feedback loops at once:
-
Editorial curation of
src/lib/models.ts(~quarterly). The pinned list is canonical orgs only:meta-llama,Qwen,openai,mistralai,microsoft,google,deepseek-ai,ibm-granite,bigcode,01-ai,CohereForAI. Each entry carries areleaseQuarterso it's obvious what's stale. Forks, distills, and uncensored remixes belong on Hugging Face, not here. -
Live "Search Hugging Face" tab in the calculator. Calls
https://huggingface.co/api/models?search=<q>&filter=gguf&sort=downloadsdirectly from the browser. Always current; classifies each result against your VRAM budget on the fly. -
Build-time HF enrichment (Phase 2, not shipped yet). A
scripts/refresh-hf-stats.tswill fetchdownloads,likes, andlast_modifiedfor every curatedhfRepoat build, write the result intodata/hf-stats.json, and surface it on each model detail page. Entries below a download threshold will get a build warning so they get reviewed in the next editorial pass.
How to propose a new model:
- Confirm it's canonical (an official org repo, not a community quant)
- Open a PR to
src/lib/models.tswithslug,name,hfRepo,params,family,type,contextK,fp16GB,summary,releaseQuarter - If MoE, set
activeParamsto active-per-forward-pass count - CI runs
pnpm buildto validate the static export still emits cleanly
Every page on vrambudget ships a parallel agent-readable view at <route>.agent (e.g. /gpu/rtx-4090.agent). The view is TOON-encoded and includes @meta, @intent, @state, @actions, @context, and @nav sections per the AVL L3 conformance spec. The full manifest lives at /agent.txt.
Built with @frontier-infra/avl.
pnpm install
pnpm dev # localhost:3000, hot reload
pnpm typecheck # tsc --noEmit
pnpm build # static export to out/, plus postbuild AVL emitter
pnpm start # serve out/ on $PORT (default 3000)out/ after build contains:
- 64 route HTMLs (home, the-math, /gpu, /model, 40 GPU detail, 20 model detail)
- 64
.agentcompanion files agent.txt,sitemap.xml,robots.txt- 61 OG image PNGs, favicon SVG, apple-touch-icon
- Static assets and font bundles
The project is wired for Railway out of the box. railway.toml pins Node 22 + pnpm 9 and configures Nixpacks with explicit build/start commands and a / healthcheck.
Works on any static host: Cloudflare Pages, Vercel, Netlify, S3+CloudFront, GitHub Pages, plain nginx. Just point at out/.
- Next.js 15 (App Router, static export)
- TypeScript 5
- React 18
- pnpm 9
- Vanilla CSS · Geist + JetBrains Mono
- @frontier-infra/avl for the parallel agent views
- Playwright for end-to-end browser verification
- Phase 2: Runtime badges. Per-model compatibility badges for Ollama, LM Studio, vLLM, and MLX with one-click install commands.
- Build-time HF enrichment. Pull download counts, license, and architecture from the HF Hub API at build so curated cards show context beyond what the slug carries.
- Per-model architecture. Replace the 13B reference for KV cache with per-model GQA / MQA + head dimensions for tighter estimates.
- Tokens-per-second. Bandwidth-based throughput estimates next to the fit pills.
- Cloud comparison. This model on your machine vs OpenRouter pricing for the equivalent.
See issues for the live list.
Open a PR. Two soft rules:
- No em dashes in new copy. Colons, periods, semicolons.
- Numbers in monospace. VRAM, params, GB, percentages.
Built by Jason Brashear. MSP owner, AI application developer, and 30-year IT veteran out of Texas.
- Personal site: jasonbrashear.com
- Substack: jasonbrashear.substack.com
- Medium: @jason_81067
- Dev.to: @webdevtodayjason
- X: @argentAIOS
- GitHub: @webdevtodayjason
MIT. See LICENSE. Use it however you want.