China Merchants Lion Rock Lab releases Lion-ER-35B-A3B, an open-weight embodied reasoning model with 3B active parameter

The Lion Rock Artificial Intelligence Lab, part of the China Merchants Institute of Advanced Technology, has released Lion-ER-35B-A3B, an open-weight embodied reasoning vision-language model. It uses a mixture-of-experts (MoE) architecture with about 35B total parameters and about 3B activated per token. The announcement was reported by QbitAI (量子位) in a Chinese-language article; the model is branded 驭物™ Lion-ER-35B-A3B in Chinese.

According to the lab, the model improves on its base across all ten public embodied-related benchmarks tested, and exceeds some larger mainstream models on several core capabilities. Context window and API pricing are not yet disclosed. Weights are published on Hugging Face (cmriat/lion-ER-35B-A3B) in BF16 Safetensors format, with native compatibility with Transformers, vLLM and SGLang.

Benchmark results

Evaluation used an avg@4 protocol: each question is answered independently four times and the scores are averaged. The lab reports:

• Equal-weighted average across ten benchmarks: 65.29
• Where2Place (spatial placement): 64.28, up 9.30 points over the base
• ERQA (embodied reasoning): 64.38, up 6.19 points over the base
• ERQA trajectory understanding subcategory: 68.18, up 12.12 points over the base
• SAT (spatial reasoning): 89.33

The base model is Qwen3.6-35B-A3B. The lab also compared against Qwen3.6-27B on four public benchmark questions, each answered four times. They covered camera-motion direction, trajectory selection, cross-view spatial reasoning and task-progress judgment. On these examples Lion-ER-35B-A3B answered all four attempts correctly, while the base and Qwen3.6-27B made repeated errors. The lab presents these as illustrative cases, not aggregate results.

Capability-driven data engine

The lab says the gain comes from data design rather than parameter count or raw data volume. Its approach has four parts:

• Capability blueprint: embodied reasoning is split into 8 top-level domains (entity, space, time, physics, affordance, task, social, evidence) and 39 second-level capabilities. Primary and auxiliary capabilities are recorded separately for composite tasks.
• Capability compilation: each target capability is turned into specific scenes, observation setups, questions and answer grounds. Samples with insufficient visual evidence are excluded.
• Verification routes: data purification (fixing image order, question references and answer mapping), geometry-based question construction (using reliable pose, depth and object correspondences), and targeted synthesis. Per the lab, answers from generating models are never treated directly as ground truth. Samples the model keeps failing go to human review.
• Eval-Driven Data Harness: evaluation failures are checked for annotation or scoring errors, then classified as one-off mistakes or systematic gaps. Confirmed gaps trigger retrieval of new material with the same reasoning logic in different scenes from a pool covering 31 data sources. The new samples are verified, added to training and re-evaluated.

The supervised training set contains 312,380 samples. Training budget is allocated by capability need rather than by the question count of each data source.

Planned work

The lab says it will next extend training data toward embodied agents: tool calling, multi-turn interaction and execution feedback, plus continuous decision-making and failure recovery in long-horizon tasks.

What this means

The reported results are specific to embodied perception and reasoning tasks: camera motion, 3D spatial relations, trajectory choice and task progress. Because the model activates about 3B parameters per token, the lab positions it as gaining these capabilities without a large rise in inference cost. All benchmark figures come from the lab's own avg@4 evaluation, and the cross-model comparisons are limited to the base model and Qwen3.6-27B. With weights open and supported by common inference frameworks, developers can test the model on their own images, video and tasks, including spatial relations, scene state, task progress and spatial constraints. The lab also describes the data-engine method as reusable.