QbitAI reported on Nov. 4 that a lab under 招商局先进技术研究院 released Lion-ER-35B-A3B, a mixture-of-experts visual-language model for embodied reasoning. The report gives it about 35 billion total parameters, with around 3 billion activated per token.
The report says the supervised training set contains 312,380 samples.
In the lab’s evaluation using four answers per question (avg@4), the model scored an equal-weight average of 65.29 across ten public benchmarks; those are lab-reported results.
QbitAI says the model weights are available on Hugging Face in BF16 Safetensors and are compatible with Transformers, vLLM and SGLang.
Source: QbitAI ↗