Will It Run? API v1 FREE
Read-only JSON, no key, CORS enabled. Same calculation code as the website, so numbers match the checker and What fits my GPU? exactly. Base URL: https://willitrunllm.com/api/v1
Endpoints
| GET | Params | Returns |
|---|---|---|
/models | old=1 include older models | slug, name, HF repo, params, MoE, release date, max context |
/hardware | type=gpu|mac|unified-apu, q= search | id, name, memory, GPU-usable memory |
/benchmarks | hw, model, quant, limit, offset | measured llama.cpp runs, each with source URL |
/fit | hw (comma-separated for multi-GPU), model, q (omit = best quant), ctx, kv (f16|q8_0|q4_0), ram, rambw, minq | verdict, memory needed, headroom, KV cache, offload plan, speed (measured or ESTIMATE) |
/fits | hw, ctx, kv, ram, minq (all|q3|q4|q5|q8, default q3), old=1 | every model that fits, best quant each |
Every response has meta.rev (data revision) and meta.dataDate. Anything computed is labelled "kind": "ESTIMATE"; real runs are "kind": "measured" with a source link.
Examples
curl
curl "https://willitrunllm.com/api/v1/fit?hw=rtx-3090,rtx-3090&model=llama-3-3-70b&ctx=8192&kv=q8_0"
curl "https://willitrunllm.com/api/v1/fits?hw=apple-m4-max-64gb&ctx=16384&minq=q4"
curl "https://willitrunllm.com/api/v1/benchmarks?hw=rtx-4090&model=gpt-oss-20b"JavaScript
const r = await fetch("https://willitrunllm.com/api/v1/fit?hw=rtx-4090&model=gpt-oss-20b&ctx=32768");
const j = await r.json();
console.log(j.verdict, j.bestQuant, j.memory.needGB, j.speed);Python
import requests
j = requests.get("https://willitrunllm.com/api/v1/fits",
params={"hw": "rtx-3090,rtx-3090", "ctx": 8192}).json()
for m in j["models"][:10]:
print(m["name"], m["bestQuant"], m["memory"]["needGB"], m["verdict"])Limits
Free, no key. Please keep it under 120 requests/minute per IP (enforced on a best-effort basis) and cache responses: data changes at most daily, and responses are cached at the CDN for an hour. Need keys, higher limits or HF GGUF lookups via the API? Join the waitlist.
Terms & license
- Free for any use, including commercial. Provided as-is with no warranty; memory and speed estimates can be wrong.
- Attribution requested: please link to willitrunllm.com ("Data from Will It Run?") where you show the data.
- Our compilation, fit calculations and estimates are licensed CC BY 4.0.
- Each benchmark row links its source. Sources: llama.cpp GitHub discussions (posts by community members under GitHub's terms; no separate license), LocalScore public results (the LocalScore code is Apache-2.0; the results site states no data license), lhl/strix-halo-testing (Apache-2.0), strix-halo-guide (MIT), one zenn.dev article (no license stated), reviewed community submissions, and Hugging Face model metadata (model weights have their own licenses; we only use metadata). Individual measurements are reported facts; where a source's own terms apply, they apply to that row, and CC BY 4.0 covers only our part.
- We may change or rate-limit the free API; v1 fields won't be removed without notice here.
Want to add measured data? Run the auto-benchmark.