Reddit Wants Knowledge-First LLMs, Not Agents

Pure knowledge is not dead, but it has been crowded out by retrieval tooling and agent benchmarks. Tulu 3 is one of the few open efforts that still treats broad factual coverage as a first-class goal.

• –The thread captures a real split in the market: agentic behavior is easier to demo, but broad knowledge is what many users actually miss.
• –Ai2 positions Tulu 3 around knowledge, reasoning, math, coding, and safety, which makes it a cleaner fit for this complaint than coding-first models.
• –Several commenters land on the same compromise: pair a smaller model with RAG, search, or a Wikipedia fetch layer instead of waiting for a magically omniscient base model.
• –Bigger open models like Qwen3.5 still get name-checked because raw parameter count remains one of the few reliable levers for factual breadth offline.