Orca 2
Microsoft's explanation-tuned small models that reason like models 5-10x larger.
About
Orca 2 is a pair of small language models from Microsoft Research, released November 21, 2023 in 7B and 13B sizes built on Llama 2. Where the original Orca imitated GPT-4's reasoning traces verbatim, Orca 2 introduced 'explanation tuning': teaching the small model to select the right reasoning strategy per task — step-by-step for math, direct answers for simple facts, recall-then-generate for knowledge questions — guided by cautious system prompts.
The results were remarkable for the era: Orca-2-13B matched or beat Llama-2-Chat-70B on zero-shot reasoning benchmarks including BigBench-Hard, AGIEval, and GSM8K, showing that better training signals could let small models punch far above their weight.
An honest caveat: both models were released under the Microsoft Research License, restricted to non-commercial research use only. The weights are downloadable and runnable locally, but commercial deployment is not permitted.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (metrics tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 73; Trending score 64/100 from researcher assessment.