NVIDIA RULER repository README results table (Claimed Length vs Effective Length, updated)

The living results table on the repo main branch, distinct from the frozen paper tables. Read 2026-10-04. Columns: model, Claimed Length, Effective Length, then average score at 4K/8K/16K/32K/64K/128K, Avg, and two weighted averages. Effective Length is defined in the paper as the longest tested length (4K/8K/16K/32K/64K/128K) at which the 13-task average clears the threshold equal to Llama-2-7B-chat's score at 4K (85.6). Entries marked * were reported by the MODEL AUTHORS in their own papers and imported by RULER, not measured by NVIDIA: Jamba-1.5-large (arXiv 2408.12570), FILM-7B (arXiv 2404.16811, YaRN applied without further training), Qwen2.5-1M (arXiv 2501.15383), Qwen3 (arXiv 2505.09388), EXAONE 4.0 (arXiv 2507.11407). Gemini-1.5-Pro, GPT-4-1106-preview and LWM results come from API endpoints rather than local weights. The README also instructs users to raise max_position_embeddings in config.json to evaluate beyond a model's own declared length, which is why the repo's own numbers can exceed config.json. Threshold definition and per-task tables: [[mkt-ruler-long-context-benchmark]] (arXiv 2404.06654v3).