LionRock Lab releases Lion-ER-35B-A3B embodied reasoner: about 35B total and 3B active parameters, 65.29 average across ten benchmarks
Synopsis
Lionrock Artificial Intelligence Laboratory has released Lion-ER-35B-A3B, a multimodal embodied-reasoning model fine-tuned from Qwen3.6-35B-A3B with roughly 35B total and 3B active parameters, aimed at scene understanding, spatial relationships and action-oriented visual questions. Across ten visual and embodied reasoning benchmarks it reports an equal-weight average of 65.29 under avg@4, the mean over four sampled responses per question, alongside a per-benchmark comparison table; the same training recipe is also adapted to the smaller Qwen3.5-4B and Qwen3.5-9B backbones. The release states the model has not been validated for safety-critical decisions or autonomous robot control, and that benchmark performance does not guarantee transfer to new environments.
Interpretation
Embodied reasoning now has an openly downloadable multimodal model: Lion-ER-35B-A3B ships under Apache 2.0 in BF16 Safetensors format for understanding scenes, spatial relationships, object interactions and action-oriented visual questions. It is an independently fine-tuned derivative of Qwen3.6-35B-A3B with modified weights, not an official Qwen release; the release notes that the upstream model is also Apache 2.0 and preserves its license file. The release provides per-benchmark comparison panels and a radar profile, and states that all benchmark figures and the scorecard use the same ten benchmarks, with RealWorldQA counted once in the equal-weight average.
The model pairs roughly 35B total parameters with roughly 3B active parameters, so only a small fraction is engaged at inference. That sparse-activation structure is inherited from the Qwen3.6-35B-A3B backbone, and the fine-tune keeps the same architecture rather than designing a new network. Under the same training recipe the release also trains Qwen3.5-4B and Qwen3.5-9B versions, showing each trained model beside its corresponding official base model with the absolute difference in percentage points.
The main comparison reports scores as avg@4, the mean over four sampled responses per question, not majority voting, pass@4 or best-of-four selection. The release separates these aggregation schemes explicitly and notes that reported scores use benchmark-specific prompts and scoring, with one response generated per request. The equal-weight average over the ten benchmarks is 65.29, and the release cautions that small score differences should not be read as statistically significant.
Perspective
The release scopes intended use to research and development in visual and embodied reasoning, and states that predictions and explanations can be incorrect and that spatial outputs should not be treated as verified geometry. The model has not been validated for safety-critical decisions or autonomous robot control, and benchmark performance does not guarantee transfer to new environments.
The release states that small score differences should not be interpreted as statistically significant and that comprehensive train-test contamination freedom has not been independently established. Training-mixture shares are computed over 312,380 supervised training examples rather than unique questions or tokens, so a question with several retained reasoning trajectories is counted more than once.
