Tagged: inference-speed

• 

  I Tried to Match Our Anthropic-and-OpenAI Models with Open Weights. The Pipeline Mattered More.

  For a month I tried to match our proprietary preset — a combination of Anthropic and OpenAI models — with cheap open-weight ones: GLM, Kimi, DeepSeek, Qwen. The end product came out nearly the same, because the pipeline breaks work into tiny tasks. What didn't come out the same was price and speed — and the cheap tokens kept getting eaten by retries, leaving a real-world 5x gap.

  Read more →