ModelCards
← All benchmarks

MIRACL-VISION, image

We haven't written a description for this benchmark yet; see the card notes below for how each maker ran it.

Your workloadRAG over docs · 12K in · 600 out · 50K requests/mo · 30% cached
  1. 1NVIDIA Nemotron 3.5 Embed 8B · NVIDIA
    72.64 NDCG@10
  2. 2NVIDIA Nemotron 3.5 Embed 1B · NVIDIA
    68.93 NDCG@10

“Best value”: no other model here reports a higher score at a lower cost for your workload. Scores are self-reported by each maker with different settings, so treat close calls as ties.

How each maker reported it

  • NVIDIA Nemotron 3.5 Embed 8B model card: Evaluation Dataset; Nemotron-3-Embed-8B-BF16 model card

    Visual rows from the Nemotron 3.5 Embed 8B model card; NVIDIA models at the default 1,120 px longest edge, Qwen3-VL models at max_pixels 1,843,200. ViDoRe V3 = 8 public + 2 private datasets; MIRACL-VISION = 18 languages. Text rows come from the Nemotron-3-Embed-8B-BF16 model card (column 2): the 3.5 card says its frozen text model produces identical text embeddings, so these text scores carry over. NVIDIA's Nemotron 3 Embed blog reports the 8B text model #1 on the RTEB leaderboard as of 15–16 Jul 2026.

  • NVIDIA Nemotron 3.5 Embed 1B model card: Evaluation Dataset; Nemotron-3-Embed-1B-BF16 model card

    Visual rows from the Nemotron 3.5 Embed 1B model card; NVIDIA models at the default 1,120 px longest edge, Qwen3-VL models at max_pixels 1,843,200. ViDoRe V3 = 8 public + 2 private datasets; MIRACL-VISION = 18 languages. Text rows come from the Nemotron-3-Embed-1B-BF16 model card (column 2): the 3.5 1B card says it shares that model's text embedding space (frozen language model) but does not publish its own text scores.

Only each model's own maker's number is shown; scores a maker reports for rival models are left out.