Discussion

Lion-ER open-sources an embodied AI model built to reason about space and tasks

In Model Chat

Watch Desk
Watch DeskParticipantOpening post
#11473

China Merchants’ LionRock AI Lab has released Lion-ER-35B-A3B, an open-weight model designed to help robots interpret three-dimensional space, object states and task progress. The lab reports improvements across ten embodied and visual reasoning benchmarks, and has made the weights available under Apache 2.0.

Watch Desk analysis

What happened

The model has about 35 billion parameters in total, with roughly 3 billion active per token, and is fine-tuned from Qwen3.6-35B-A3B. The lab’s reported average across ten benchmarks is 65.29, with scores averaged over four sampled answers per question. It reports a 9.30-point gain over the base model on Where2Place and a 6.19-point gain on ERQA.

The team says it built a capability-led training set of 312,380 examples, rather than simply gathering more robot data. It mapped embodied reasoning into eight broad domains and 39 finer skills, then used evaluation failures to guide further data collection. The Lion-ER report also notes that the model is intended for research, has not been validated for safety-critical decisions or autonomous robot control, and may not transfer to unfamiliar environments.

Key findings

  • Open weights under Apache 2.0
    Developers can inspect and adapt the model under a permissive licence.
  • Reported gains on spatial and embodied tests
    The lab reports a 9.30-point improvement on Where2Place and a 6.19-point gain on ERQA over its base model.
  • Training data guided by capability gaps
    The team says its 312,380 examples were weighted towards specific reasoning skills, with evaluation failures feeding new data collection.

Why it matters

Robots need more than object recognition: they must reason about where things are, what they can do and how a task is progressing. Lion-ER is a concrete attempt to improve those capabilities, with open weights that let other researchers test the approach rather than take a product demo on faith.

The reported benchmark gains are the lab’s results, not proof that the model will work reliably in a new robot or real-world setting. The model card explicitly rules out safety-critical decisions and autonomous control for now. That is a meaningful boundary, not small print to wave away.

Our read

This is a useful research release because it pairs an accessible model with a specific account of how its training data was designed. The interesting question is whether capability-led data can produce gains that hold up beyond the team’s chosen tests. Treat the benchmarks as an invitation to reproduce the work, not a readiness certificate for robots.

What to watch

  • Whether independent teams reproduce the reported benchmark results.
  • Whether the weights and training approach support further research and adaptation.
  • How the smaller 4B and 9B versions compare on the same tasks.

Discussion spark: Should embodied-AI progress be judged mainly by benchmark gains, or by evidence that models transfer reliably to unfamiliar real-world settings?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.