Reactions from ranked influencers
2 posts
Added to prinzbench: Kimi K3. K3 is by far the best open-source model I have tested to date. Its overall score (47/99) is significantly higher than GLM-5.2's (30/99). For context, 47/99 is slightly lower than the score achieved by Gemini 3.1 Pro in February (50/99). I view K3 as being roughly in the same class as Gemini 3.1 Pro - albeit a bit better at search and a bit worse at logical reasoning. K3 is also *significantly* more inconsistent than Gemini 3.1 Pro and similar models. At times, K3 struggled with fairly easy questions that most other modern models are able to nearly always answer correctly; but, on the other hand, some of K3's responses to much harder questions were really great. K3 is the first non-OpenAI model to have answered 10+ Search questions correctly on my benchmark (its Search sub-score was 11/24). This is really impressive, and I would recommend K3 for web search tasks without hesitation to anyone. It's a good model, sir! Just don't expect it to be anywhere close to the frontier.
Reminder that Prinzbench is about the first party search implementation than about model capability, which is why Grok 4.20 is so powerful relative to its standing in normal LLM tasks. Kimi service is really good
Added to prinzbench: Kimi K3. K3 is by far the best open-source model I have tested to date. Its overall score (47/99) is significantly higher than GLM-5.2's (30/99). For context, 47/99 is slightly lower than the score achieved by Gemini 3.1 Pro in February (50/99). I view K3 as being roughly in the same class as Gemini 3.1 Pro - albeit a bit better at search and a bit worse at logical reasoning. K3 is also *significantly* more inconsistent than Gemini 3.1 Pro and similar models. At times, K3 struggled with fairly easy questions that most other modern models are able to nearly always answer correctly; but, on the other hand, some of K3's responses to much harder questions were really great. K3 is the first non-OpenAI model to have answered 10+ Search questions correctly on my benchmark (its Search sub-score was 11/24). This is really impressive, and I would recommend K3 for web search tasks without hesitation to anyone. It's a good model, sir! Just don't expect it to be anywhere close to the frontier.