The Lion Mountain AI Lab under China Merchants Advanced Technology Institute has released Lion-ER-35B-A3B, an embodied-reasoning vision-language model built on a mixture-of-experts architecture with roughly 35B total parameters and only about 3B activated per token. The lab claims across-the-board gains on ten embodied-AI benchmarks, with some capabilities surpassing larger models, and attributes this to a capability-first data engine that decomposes embodied reasoning into 8 top-level domains and 39 sub-capabilities and iterates data against evaluation gaps rather than scaling raw data volume. No concrete benchmark numbers, baseline configurations, or release/access details are disclosed, so the claims remain vendor assertions.
Vision-Language Model (VLM)
视觉语言模型
Multimodal AI models capable of processing and understanding both image and text data simultaneously.
Curated Intelligence
A new paper from Google DeepMind and Stanford reveals that vision-language models are susceptible to hidden steering by irrelevant geometric shapes in prompts. Experiments showed that adding a red heart image caused Claude Opus 4.7 and GPT-5.4 to shift their recursive solution choice from 0% to 100% on binary tree tasks, while Gemini's reasoning traces often failed to flag the visual influence. This highlights the critical need for rigorous isolation testing of multimodal inputs.
OXR releases a 1.3B-parameter open-source document parser that pioneers Recursive Self-Improvement (RSI) in OCR training. By using VLMs to evaluate outputs and cluster failure causes for targeted data generation, the model tops olmOCR-Bench, OmniDocBench, and DocParsingBench while achieving 3.05x the throughput of GLM-OCR on a single RTX PRO 6000.
Micron is developing Edge Inference Memory (EIM), a specialized storage chip for autonomous driving and humanoid robots that stacks low-power memory to reduce latency and boost bandwidth. Distinct from Processing-in-Memory (PIM), EIM aims to integrate with LPDDR6 by late 2028 to meet the growing demands of physical AI systems.
Gemini 4 Argon, within the Antigravity framework, successfully generated a complex Japanese temple structure in Blender using a single prompt containing detailed workflow instructions. The process demonstrates a full automation loop: blueprint planning, reference image generation via Nano Banana, multi-agent collaborative modeling, and iterative refinement through visual comparison between Blender renders and references. A coordinator agent managed tasks while a critic agent validated outputs, ensuring accuracy against visual targets. The final result extended to procedurally generated animations and ambient sounds, highlighting significant advancements in LLMs' ability to understand real-world physics and execute long-horizon complex tasks.
ModelBest and OpenBMB have open-sourced MiniCPM-V 4.7, a multimodal LLM featuring a 35B-parameter MoE architecture that activates only 3B parameters per inference. The model runs offline on devices like the iPhone 17 Pro Max using just ~3GB of RAM, achieving decoding speeds exceeding 13 tokens per second after optimization. It also demonstrates strong performance in general vision, math, and video understanding benchmarks, with broad support for mobile OSes and major chip platforms.
DeepSeek plans to lead a multi-million dollar seed round for Winverse AI, founded by ex-Tencent scientist Han Hu, at a $400 million valuation. This move extends DeepSeek's robotics strategy from hardware (Unitree) to embodied intelligence models, aiming to build the 'brain' for robots.
Liquid AI has released and open-sourced the d1-3B and d1-omni-600M decision models, emphasizing ultra-low latency and multimodal capabilities. The d1-3B model achieves 8ms text inference on an RTX 4090 and leads in Decision Index v0.2.1 benchmarks among sub-10B models. The d1-omni-600M variant adds voice input support, offering a lightweight solution for real-time agent applications on edge devices.
Liquid AI released the experimental draft model LFM2.5-VL-DSpark for its vision-language model LFM2.5-VL-3B. By leveraging speculative decoding, it achieves up to 3.1x faster inference with minimal memory overhead while maintaining output quality.
NVIDIA announced official support for GPU hardware video decoding via PyNvVideoCodec within vLLM, addressing the throughput bottleneck caused by CPU-based decoding in multi-GPU video captioning tasks. By offloading frame decoding from CPU to GPU, this integration significantly enhances end-to-end inference efficiency and scalability for Vision Language Models (VLMs) handling short-output video annotation scenarios.