BEAM (128K)
General QA · 2025-10-31
The 128K-token tier of BEAM (Beyond a Million Tokens), a long-term memory benchmark of long multi-session conversations across domains such as coding and math. With the whole conversation in context, the model answers memory-probing questions covering ten abilities such as factual recall, temporal reasoning and state tracking, and an LLM judge scores each answer against a reference.
Top models (higher is better)
| Model | Score |
|---|---|
| Muse Glimmer 30B | 66.2 |
| Qwen3.6 27B | 63.0 |
| Gemma 4 31B IT | 58.2 |