#post-training

LLM X/Twitter Apr 23, 2026 1 min read

Perplexity says Qwen post-training beats GPT on factuality cost

Why it matters: search products need factuality and citations, not just fluent answers. Perplexity said its SFT + RL pipeline lets Qwen models match or beat GPT models on factuality at lower cost.

#perplexity #qwen #retrieval

LLM Apr 16, 2026 2 min read

Lightning OPD cuts reasoning-model post-training to 30 GPU hours

Lightning OPD attacks a practical bottleneck in on-policy distillation: keeping a live teacher model running throughout training. The paper reports 69.9% on AIME 2024 from Qwen3-8B-Base in 30 GPU hours, a 4.0x speedup over standard OPD.

#llm #distillation #post-training