Pokee-Isaac 28B Report - MRCR v2 8-needle 256K: leaderboard
OpenAI MRCR v2 multi-round coreference, 8-needle split: eight near-identical requests and their responses are hidden in a long synthetic conversation in the 256K-token bin, and the model must reproduce the response to the specified instance; 30 paired samples, official random-prefix filter and SequenceMatcher score on a 0-1 scale. Models with no usable output at this length are left out. From Pokee AI's Pokee-Isaac 28B technical report (September 10, 2026 revision): Pokee ran every listed model itself through hosted APIs under one harness, with each peer's default temperature (unless a benchmark section fixes one) and a high thinking level where available; exhausted retries, context overflows and verifier failures count as zero. Higher is better.
Metric: Score (0-1). Source: pokee.ai. Saturation forecast: Around June 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.5 Flash Lite | 0.47 | #300 |
| 2 | GPT-5.6 Luna | 0.21 | #117 |
| 3 | Nemotron 3 Super | 0.14 | #766 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=pokee-isaac-28b-report-mrcr-v2-8-needle-256k · How It Works · Data refreshed daily, snapshot 2026-11-12.