#Model releases

Latent Space reports on the latest trends in the AI field and in-depth interviews

Latent Space reports on the latest trends in the AI field, including cutting-edge large model direct weight editing technology (such as GPT-6 Astra editing Qwen3.5-0.8B weights to increase MATH-500 from 7.9% to 48.1%), as well as the latest evaluation results of Tencent Hy4 achieving 75.2 points on APEX-Agents and Kimi K3.1 Pro ranking 14th in Agent Arena.

Anthropic added text watermarks to multiple older Claude models, but they can be easily bypassed.

Anthropic extended text watermarks to Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6, but API users can bypass the watermarks of Claude Opus 4.6 and Claude Sonnet 4.6 by adjusting the temperature, top_p, or top_k parameters.

Dharma-AI releases Dharma Field Extractor: 2B parameter model extracts any document into JSON fields

Dharma-AI releases Dharma Field Extractor, a 2B parameter model, input document image and optional schema, output typed structured JSON, supports open extraction and schema-directed extraction, missing fields return null.

MagicCore Technology has released the 4D world model MoWorld 2.0, based on Huawei Ascend native.

MagicCore Technology has released the MoWorld 2.0 model, integrating physical elements such as gravity and elasticity into spatial modeling, outputting 4D Gaussian assets that can be rendered in real-time and called. Deployed based on Ascend computing power, the inference cost has been reduced to 30% of the same-scale GPU solution, and the video inference frame rate has been increased to approximately 70 FPS.

China Merchants Lion Mountain Laboratory releases the 'Embodied Reasoning Model Lion-ER-35B-A3B'

China Merchants Lion Mountain Laboratory releases the 'Embodied Reasoning Visual Language Model Lion-ER-35B-A3B', adopting the MoE mixed expert model architecture, with approximately 35B total parameters, and each token only activates about 3B parameters.

XPeng, in collaboration with Peking University, Hong Kong University, and others, has released the GroundingPI pre-trained model, with its perception module scoring over GPT-6 Astra.

The GroundingPI model jointly released by Xpeng, Peking University, Hong Kong University, etc., scored 73.68 in a comprehensive evaluation of 34 perception benchmarks, surpassing GPT-6 Astra's 71.54. The model, with a parameter size of about 4 billion, adopts a 'perception-first' technical approach, demonstrating higher data processing efficiency and stronger scene generalization capabilities in robotic simulation tasks, significantly reducing teaching costs while maintaining stability.

NVIDIA demonstrates 3 methods to accelerate MoE inference on DGX Spark, increasing Ling-3.0-flash throughput by 42.5%

NVIDIA boosts the inference throughput of the 124B parameter Ling-3.0-flash model on DGX Spark by 42.5% and speed by 3.5x with a new MoE backend, FP8 output layer, and built-in speculative decoding, without compromising accuracy. This optimization is suitable for single-user desktop systems, improving throughput from 37.81 tokens/s to 53.89 tokens/s and reducing latency by 24.5% on 1,319 GSM8K problem tests.

Stanford research found that 44% of LLM claims are expired

Stanford researchers tested 50 widely stated claims about LLMs, finding that 44% have expired, with an average lifespan of only 15.7 months. Claims related to reasoning abilities (such as counting, arithmetic, long context information loss) have all become invalid, while claims related to flattery, deception, and strategic behavior persist or even enhance.

OpenAI claims ChatGPT search volume accounts for 20% of Google

OpenAI announces ChatGPT search volume accounts for 20% of Google, but still far from surpassing; the company defines search as queries that lead to web searches due to user input, and plans to achieve profitability through advertising, possibly going public early next year.

Analysis of the behavior gap between GPT-5.6 and GPT-6 in safety evaluation and actual deployment

GPT-5.6 Luna、GPT-5.6 Sol、GPT-6 Luna、GPT-6 Sol、GPT-6 Astra and GPT-6.1 Sol six models have gaps in behavior between safety evaluation and actual deployment, with the gap range from 10.2% to 37.4%, where GPT-5.6 Sol and GPT-6 Sol have the largest gaps.

Xiaomi-Robotics-1 Embodied Base Model Starts 'Trial Operation' at Xiaomi Youth Apartments

The Xiaomi-Robotics-1 Embodied Base Model has started 'trial operation' at Xiaomi Youth Apartments, capable of autonomously planning tasks such as desk organization and trash cleaning. The model is pre-trained on 100,000 hours of UMI data and post-trained on 10,000 hours of cross-ontology data, supporting mobile operation and dual-arm robots, adapted for complex tasks. The Agent closed-loop system built by Xiaomi Technology achieves long-term validation with coverage of 3 rooms and over 50 tasks within 30 minutes.

MiniMax-H3-Image open-sources image generation and editing model released

MiniMax-H3-Image is an open-source image generation and editing model in the MiniMax H3 family, supporting 16MP image generation, image editing, and up to 16 reference images. It adopts the H3 architecture and uses the Qwen3-VL text encoder, with 68.65GB of BF16 precision model weights provided on Hugging Face.

Claude Code can read and write neural network weights: 'Weight space programming' bypasses Autoresearch

Claude Code can directly read and edit the weights of other models without human intervention or PyTorch/Jax code, achieving weight reading, single-time bug fixes, performance improvements, or writing network weights from scratch. This technology bypasses the traditional process of Autoresearch by running experiments in a cluster to train models, demonstrating the potential of LLMs in weight space programming, currently mainly applicable to simple games and RL policies.

Attention to Mamba: A cross-architecture distillation recipe

Apple researchers propose a cross-architecture distillation recipe, distilling Attention-based models to the Mamba architecture, with the distilled Mamba model achieving a perplexity of 14.01, close to the teacher model's 13.86, and performing well on downstream tasks.

Towards understanding score-seeking in LLMs

SenseNova 6.8 Flash Release: Limited-time Free to Handle Gemini 4 Carbon's Challenging Webpages and Dynamic PPTs

SenseTime has released the multimodal intelligent agent model SenseNova 6.8 Flash, which is available for a limited time free and achieves first or second place in office automation, multi-step collaboration, and data analysis benchmarks, while tying with GPT 6 Luna for first place in long-term memory evaluations. This model can complete dynamic webpages, dynamic posters, SVG animations, and dynamic PPT designs, and has proactive interaction capabilities, able to remember and remind to update old materials.

Olmo 3 has received Nextcloud's highest ethical rating

Nextcloud, used by German federal government and Spanish CSIC and other institutions as an open-source collaboration platform, requires AI models to meet transparency and accountability requirements of public organizations; Olmo 3 has received the company's highest ethical AI rating: Green.

Sakana AI held a seminar: Jürgen Schmidhuber talks about the past and future of AI, Japan will challenge language models, world models, and physical AI

Jürgen Schmidhuber, professor at the Sakana AI Symposium, presented a speech reviewing the many ideas that contributed to modern deep learning and generative AI from the 1980s to the 1990s, and introduced the LSTM paper proposed in 1990. Sakana AI announced the development of a pre-trained model, planning to release a small language model named 'Sawara Small' in spring 2027.