Renaissance LabsLearn

Models · The open ecosystem

Inside the strongest open models

Architecture, data, recipes and what matters

14 minIntermediate18 chapters

The best open-weights model scores 46 on the Artificial Analysis Intelligence Index (v4.3.2) and the best closed model 58. The film compares ten open models: mixture of experts, long-context attention, training stability, data, and reinforcement learning. Its reading is that agentic post-training explains more of the gap than size or token count.

What to take away

  • Index scores only compare within one version, because the benchmark gets harder with each release.
  • Nearly every top open model is a sparse mixture of experts, so only a small share of weights is used per token.
  • Compressed and hybrid attention keep the KV cache manageable at a million tokens of context.
  • Active parameters and training tokens predict score only weakly across these ten models, while agentic reinforcement learning stands out.
  • On Arena, which measures human preference, only one open model reaches the top fifteen.

Go deeper

Read the transcript106 lines, click any line to jump

Sources

Where a source has no link, it is cited from the film's research notes. Estimates and unverified claims are labelled in the film.

Keep watching