model string | date timestamp[s] | suite string | items int64 | seeds int64 | max_tokens int64 | latency_requests int64 | spent_usd float64 | providers list | notes string |
|---|---|---|---|---|---|---|---|---|---|
meta-llama/llama-3.3-70b-instruct | 2026-10-07T00:00:00 | gsm8k-200 | 200 | 1 | 384 | 20 | 1.13 | [
{
"provider": "DeepInfra",
"tag": "deepinfra/turbo",
"quantization": "fp8",
"context_length": 131072,
"price_in_per_m": 0.09999999999999999,
"price_out_per_m": 0.32,
"served_by": "DeepInfra",
"accuracy_pct": 98,
"accuracy_per_seed": [
98
],
"accuracy_errors": 0,
... | Providers first run with request failures (Cloudflare rejects `seed`; rate limits elsewhere) were re-run with retries and a seed fallback; failures are reported separately. |
Same model, many providers: Llama 3.3 70B on OpenRouter
Independent measurements of the providers serving meta-llama/llama-3.3-70b-instruct on OpenRouter (2026-10-06/07): GSM8K accuracy (5-shot, numeric scorer), output throughput and latency per provider, with failures reported separately. Write-up: Lab notes #3.
llama-3.3-70b-2026-10-07.json— results as published on the blog
Measured through OpenRouter with provider pinning; results describe those dates only.
- Downloads last month
- 34