Developer API
The calculators, as JSON.
A free, static, CORS-open API. Every figure is computed by the same shared calculation engine (v1.1.0) that powers the on-site tools — the API cannot disagree with the calculators. No key, no rate limits beyond ordinary politeness, cached 24 h.
GET onlyCORS *engine v1.1.0reviewed July 2026
Endpoints
| Endpoint | Returns |
|---|---|
/api/v1/index.json | Directory, engine version, formula constants |
/api/v1/models/index.json | Slim list of all 506 model profiles |
/api/v1/models/{slug}.json | Full profile: weights/overhead/total/recommended at six quants, single-GPU fit list, award picks at Q4 |
/api/v1/gpus.json | 12-GPU catalog: VRAM, usable VRAM, bandwidth, TDP, indicative prices |
/api/v1/quants.json | Quantization table (bits per weight) |
/api/v1/providers.json | 18 cloud provider profiles with indicative H100 bands |
Example
curl https://bitpute.com/api/v1/models/llama-3-1-8b-instruct.json
{
"engine": "1.0.0",
"model": {
"name": "Llama 3.1 8B Instruct",
"paramsB": 8.03,
"estimates": {
"q4_k_m": { "weightsGB": 4.54, "totalGB": 5.51, "recommendedGB": 6.06 },
"fp16": { "weightsGB": 14.96, "totalGB": 16.45, "recommendedGB": 18.1 }
},
"fitsAtQ4": ["rtx-3060", "rtx-4060-ti", "…"],
"awardsAtQ4": { "cheapest": "rtx-3060", "inference": "rtx-3090" }
}
}
Methodology & caveats
Weights = params × bits-per-weight ÷ 8 (GiB). Runtime overhead = 0.75 GB + 5% of weights. Recommended adds 10% headroom. KV cache is context-dependent and deliberately excluded from static estimates — size it with the GPU Memory Calculator. Prices are indicative street/cloud figures, reviewed July 2026. Slugs match the URLs of /models/ and /gpu/ pages.
Attribution appreciated: link to bitpute.com where the data is shown.