Benchmark Almanac

RefSpatial-Bench

Qwen3.5 397B-A17B 73.6% leads since February 2026
Fading

Real images where the model points to an object or free space defined by multi-step spatial relations. Scored by mask hit.

No independent leaderboard. Subset aggregation not stated. One cross-checked row (RoboBrain2.0-7B) differs from its origin: 42.2 vs 32.50.

Released
June 2025
Kind
Embodied reasoning
Level
Everyday
Scored
Checked automatically
Answers
Short answer
Tasks
All public
Strongest model tested
Qwen3.5 397B-A17B #77 on the capability index
Last in a launch post
April 2026, dropped before it saturated
Last model added
May 2026
16
Models scored in lab cards
12
Launch posts list it, from 2 labs
6
System cards report it, from 4 labs

Qwen3.5 397B-A17B leads with 73.6% among models released by July 2026

Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.

The newest model with a score here came out in July 2026, and nobody has reported a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

16 models. Scroll sideways for the rest.

  1. 73.6 Qwen3.5 397B-A17B
  2. 72.6 Seed2.0 Pro
  3. 70.0 Qwen3.6 27B
  4. 69.9 Qwen3-VL-235B-A22B
  5. 69.3 Qwen3.5-122B-A10B
  6. 67.7 Qwen3.5-27B
  7. 64.3 Qwen 3.6 35B-A3B
  8. 63.5 Qwen3.5-35B-A3B
  9. 58.5 Qwen3.5-9B
  10. 56.3 Seed1.8
  11. 54.6 Qwen3.5-4B
  12. 53.1 Cosmos3-Nano
  13. 48.5 gemini-robotics-er-1.5-preview
  14. 48.4 Cosmos3-Edge
  15. 32.9 Qwen3.5-2B
  16. 23.5 Qwen3.5-0.8B
  1. 1 Qwen3.5 397B-A17B 73.6%
  2. 2 Seed2.0 Pro 72.6%
  3. 3 Qwen3.6 27B 70%
  4. 4 Qwen3-VL-235B-A22B 69.9%
  5. 5 Qwen3.5-122B-A10B 69.3%
  6. 6 Qwen3.5-27B 67.7%
  7. 7 Qwen 3.6 35B-A3B 64.3%
  8. 8 Qwen3.5-35B-A3B 63.5%
  9. 9 Qwen3.5-9B 58.5%
  10. 10 Seed1.8 56.3%
  11. 11 Qwen3.5-4B 54.6%
  12. 12 Cosmos3-Nano 53.1%
  13. 13 gemini-robotics-er-1.5-preview 48.5%
  14. 14 Cosmos3-Edge 48.4%
  15. 15 Qwen3.5-2B 32.9%
  16. 16 Qwen3.5-0.8B 23.5%

Source: Labs’ model cards

#ModelScoreSettingDateSource
1 Qwen3.5 397B-A17B 73.6% Feb 2026 Lab model card
2 Seed2.0 Pro 72.6% Feb 2026 Lab model card
3 Qwen3.6 27B 70% Apr 2026 Lab model card
4 Qwen3-VL-235B-A22B 69.9% Sep 2025 Lab model card
5 Qwen3.5-122B-A10B 69.3% Feb 2026 Lab model card
6 Qwen3.5-27B 67.7% Feb 2026 Lab model card
7 Qwen 3.6 35B-A3B 64.3% Apr 2026 Lab model card
8 Qwen3.5-35B-A3B 63.5% Feb 2026 Lab model card
9 Qwen3.5-9B 58.5% Mar 2026 Lab model card
10 Seed1.8 56.3% Dec 2025 Lab model card
11 Qwen3.5-4B 54.6% Mar 2026 Lab model card
12 Cosmos3-Nano 53.1% May 2026 Lab model card
13 gemini-robotics-er-1.5-preview 48.5% Sep 2025 Lab model card
14 Cosmos3-Edge 48.4% May 2026 Lab model card
15 Qwen3.5-2B 32.9% Mar 2026 Lab model card
16 Qwen3.5-0.8B 23.5% Mar 2026 Lab model card

3 models have held number one in labs’ own numbers since its release in June 2025

Lab claims count from the day the card came out, independent runs from the model’s release.

Qwen ByteDance Seed
Lab claims Not tested since May 2026
Qwen3.5 397B-A17B 73.6%

Source: Labs’ model cards

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Qwen3.5 397B-A17B Qwen3-VL-235B-A22B 0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Qwen3.5 397B-A17B

Source: Labs’ model cards

Labs headline it less often now, last in April 2026

Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.

Qwen
ByteDance Seed
DateLaunch postScoreAs the post writes it
Apr 22, 2026 Qwen3.6 27B 70 RefSpatialBench
Apr 15, 2026 Qwen 3.6 35B-A3B 64.3 RefSpatialBench
Mar 2, 2026 Qwen3.5-0.8B 23.5 RefSpatialBench 23.5/21.7 (with CI / without CI)
Mar 2, 2026 Qwen3.5-2B 32.9 RefSpatialBench 32.9/30.0 (with CI / without CI)
Mar 2, 2026 Qwen3.5-4B 54.6 RefSpatialBench
Mar 2, 2026 Qwen3.5-9B 58.5 RefSpatialBench
Feb 24, 2026 Qwen3.5-35B-A3B 63.5 RefSpatialBench
Feb 24, 2026 Qwen3.5-27B 67.7 RefSpatialBench
Feb 24, 2026 Qwen3.5-122B-A10B 69.3 RefSpatialBench
Feb 16, 2026 Qwen3.5 397B-A17B 73.6 RefSpatialBench
Dec 18, 2025 Seed1.8 56.3 RefSpatialBench
Sep 23, 2025 Qwen3-VL-235B-A22B 69.9 RefSpatialBench

6 system cards report it, from 4 labs

Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.

DateDocumentScoreAs the document writes itSection
May 31, 2026 Cosmos3-Nano technical report 53.1 RefSpatial, Cosmos3-Nano 8B column, Table 10 Capabilities
May 31, 2026 Cosmos3-Edge technical report 48.4 RefSpatial, Cosmos3-Edge 2B column, Table 10 Capabilities
Mar 21, 2026 Seed1.8 model card 56.3 RefSpatialBench (Table 2) Capabilities
Feb 14, 2026 Seed2.0 Pro model card 72.6 RefSpatialBench (Table 8) Capabilities
Nov 26, 2025 Qwen3-VL-235B-A22B technical report 69.9 RefSpatialBench (Table 2) Capabilities
Sep 25, 2025 gemini-robotics-er-1.5-preview technical report 48.5 RefSpatial (Table 19, thinking on). Thinking off 41.8 Capabilities

Versions

Replaced by RefSpatial-Expand-Bench

More in Robotics

All 103 in Robotics
  • ERQA In use 400 multiple-choice visual questions from robot and egocentric footage on spatial, trajectory, action and state reasoning. Accuracy. Gemini Robotics ER 2 78.5%
  • RoboDojo In use Generalist manipulation policies run 42 Isaac Sim tasks and 18 remote real-robot tasks. Scored by task score and success rate. GPT-6 Astra 29%
  • RoboLab In use Zero-shot DROID Franka policies in Isaac Lab on 120 tabletop tasks at three instruction levels, 10 episodes each. Success rate. FLUX 3 Action 42.9%
  • Progress classification (Google DeepMind) In use Models assign each frame of a robot video to one of five task-progress bands from 0 to 100 percent. Scored by accuracy. Gemini Robotics ER 2 57.4%
  • Human proximity (Google DeepMind robotics) In use Models judge from robot camera views whether a human is within 1 meter, a spatial awareness check for safe robot behavior. Scored by accuracy. Gemini Robotics ER 2 93%
  • Moment finding (Google DeepMind) In use Models find the moment in a robot video when an event happens. Scored by accuracy and mean absolute time distance in seconds. Gemini Robotics ER 2 91.3%