The small, fast Meta model is Llama 3.1 8B Instruct — the efficient tier at 240 tokens/sec and $0.020 per million input tokens. It trades a few points of raw intelligence for speed and cost, the right call for high-volume, latency-sensitive work. Llama 3.2 3B Instruct (free) (149 t/s) and Llama 3.2 3B Instruct (149 t/s) round out the top three.
Compact, efficient models — the small/mini/flash/haiku tier — ranked by output speed. These trade a little raw intelligence for low cost and high throughput, which is the right tradeoff for chat, classification, extraction and other high-volume work.
The small, fast Meta model is Llama 3.1 8B Instruct — the efficient tier at 240 tokens/sec and $0.020 per million input tokens. It trades a few points of raw intelligence for speed and cost, the right call for high-volume, latency-sensitive work. Llama 3.2 3B Instruct (free) (149 t/s) and Llama 3.2 3B Instruct (149 t/s) round out the top three.
Llama 3.2 3B Instruct (free) (149 t/s) is the closest alternative on this metric, followed by Llama 3.2 3B Instruct (149 t/s). See the full ranking above for the tradeoffs.
modelgrep tracks 13 Meta models with live benchmarks, speed, latency and per-provider pricing, led on intelligence by Llama 4 Maverick. 9 of them qualify for this ranking.