NaiveAI Promises 2,100 Tokens per Second. How Well Does It Code?
With thinking enabled, NaiveAI built competitive apps, but its fastest output did not guarantee faster results on hard problems.
Latent Space rated its solar system app best, its tree competitive with DeepSeek’s, and its working ray tracer less efficient than DeepSeek’s. With thinking disabled, it failed to produce a usable tree app. Across 30 questions, median response time was 3.1 seconds, but long reasoning traces erased much of the average speed advantage; answer accuracy was not graded. The advertised 2,100 tokens per second was a one-second peak, and throughput measurements used different sources across models.
Latent Space's note
Latent Space rated its solar system app best, its tree competitive with DeepSeek’s, and its working ray tracer less efficient than DeepSeek’s. With thinking disabled, it failed to produce a usable tree app. Across 30 questions, median response time was 3.1 seconds, but long reasoning traces erased much of the average speed advantage; answer accuracy was not graded. The advertised 2,100 tokens per second was a one-second peak, and throughput measurements used different sources across models.
Latent Space's note
score 5