MemLens (32K): leaderboard

MemLens tests the multimodal long-term memory of vision-language models with 789 questions whose answers depend on images and facts spread over many sessions. Overall accuracy in percent at a 32K-token interleaved image and text conversation history, with the evidence sessions hidden among distractor sessions (MemLens, 2026), judged by Qwen3-VL-235B; higher is better.

Metric: Overall accuracy (%) on all 789 questions (information extraction, multi-session, temporal and knowledge-update reasoning and answer refusal) at a 32K-token history, interleaved multimodal multi-session conversation history with the evidence sessions hidden among haystack sessions, judged correct or incorrect by a Qwen3-VL-235B-A22B-Instruct judge; higher is better. Source: arxiv.org. Saturation forecast: Around June 2028. 26 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3.5 122B A10B58.68#346
2Qwen 3 VL 30B A3B Instruct55.01#895
3Kimi K2.554.88#266
4Qwen 3 VL 235B A22B Instruct53.23#532
5Qwen 3.5 27B52.98#375
6GPT-5.452.72#135
7GLM-4.6V52.34#710
8Qwen 3.5 9B50.32#828
9Qwen 3 VL 8B Instruct49.18#1029
10Qwen 3 VL 235B A22B (Thinking)48.29#459
11Cosmos-Reason2-8B46.13#755
12Qwen 3 VL 8B (Thinking)46.01#910
13Nemotron Nano 12B v2 VL44.99#2368
14Qwen 3.5 4B44.36#1241
15Qwen 3.5 2B44.36#5307

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=memlens-32k · How It Works · Data refreshed daily, snapshot 2026-11-21.