China Merchants' Lion Mountain Lab open-sources Lion-ER-35B-A3B, a 35B MoE embodied-reasoning VLM activating ~3B per token

On November 4, QbitAI reported that the Lion Mountain AI Lab, part of China Merchants Group's Advanced Technology Institute, released Lion-ER-35B-A3B, a vision-language model built for embodied reasoning, and published its weights on Hugging Face. The model uses a mixture-of-experts architecture with about 35B total parameters and only around 3B activated per token.

The problem the lab targets is not object recognition but the gap between what a camera shows and what is physically true. An object that sits higher in the frame is not necessarily higher in the room; furniture looks different from a new angle but remains the same object; apparent movement in a video may come from the camera turning rather than the object moving. Confusing these cases leads directly to failed grasps, wrong routes and premature task steps, so the model has to reason about 3D relations, track state changes and judge which stage a task has reached.

Rather than scaling up parameters, the team focused on how training data is designed. The article notes that efforts such as Open X-Embodiment and DROID have kept expanding real-world robot data, yet the lesson of that expansion is that more data does not automatically mean more capability, and prior work has shown that merely reweighting data domains can shift downstream results.

The first element is a capability blueprint that splits embodied reasoning into eight top-level domains — entity, space, time, physics, affordance, task, society and evidence — and 39 sub-capabilities, with reasoning style, input format and answer format annotated as separate dimensions to avoid mixing taxonomies. The blueprint specifies not only what the model should learn but under which conditions each capability must be invoked, and composite tasks record primary and auxiliary capabilities separately.

The second element is what the team calls capability compilation: turning each abstract capability into concrete training items by deciding what scene is needed, how observations are organized, what question is asked and what evidence fixes the answer. Three supervision routes — data purification, geometry-grounded question construction and targeted synthesis — keep each item anchored to evidence. The stated principle is to keep hard problems that have evidence and exclude pseudo-hard ones that do not; a model failure is not a reason to delete a sample, and agreement between models is not proof of correctness, while persistently failed hard cases go to human review.

The resulting supervised set contains 312,380 samples, with training budget allocated by capability need rather than by how many questions a given data source happens to contain. Evaluation is then wired into data production through an Eval-Driven Data Harness: errors are checked for labelling, visual evidence and scoring problems, gaps are confirmed, and new material with the same reasoning logic but a different scene is retrieved from a pool covering 31 data sources and roughly 4.3 million records before entering the next round.

On the ten public benchmarks the lab reports, the model reaches an equal-weight average of 65.29 under an avg@4 protocol, meaning each question is answered four times independently and averaged. Where2Place scores 64.28, up 9.30 over the base model; ERQA scores 64.38, up 6.19; the SAT spatial reasoning benchmark reaches 89.33; and the ERQA trajectory-understanding subclass reaches 68.18, a gain of 12.12. The lab says every one of the ten benchmarks improved over its base model, and its worked examples also compare against Qwen3.6-27B.

The four published examples map onto camera-motion judgement, trajectory selection, cross-view spatial mapping and task-progress recognition: deciding whether a bottle shifted because the camera tilted down, picking the one trajectory out of four that can actually reach a charger before moving it, reconstructing where a painting sat relative to the photographer using photos from another room, and separating a robot arm that is approaching a sushi roll from one that has arrived. These are the points where a language instruction becomes a physical action and where a single misjudgement derails the whole sequence.

Lion-ER-35B-A3B ships as BF16 Safetensors weights and works natively with Transformers, vLLM and SGLang, so developers can evaluate and adapt it on their own images, video and tasks. The lab says the next step is extending the same data engine toward embodied agents, covering tool calling, multi-turn interaction and execution feedback, and exploring continuous decision-making and failure recovery in long-horizon tasks.