The open-source vector database Chroma has released Context-1, a 20 billion parameter search model for multi-turn retrieval. This model utilizes self-editing context technology, trained on multiple tasks through reinforcement learning and a curriculum mechanism, demonstrating outstanding performance, particularly in the fields of web, finance, and law, and also showcasing cross-domain search capabilities in the email sector.

# Article Source: Axis Axis Robotics is restructuring data diversity and scaled production methods for embodied AI through a simulation-first strategy. In 2025, multiple technological pathways in the robotics industry are rapidly converging: the commoditization upgrade of embodied hardware supply chains has made it possible for expensive prototypes to achieve scaled deployment for the first time; Vision-Language-Action (VLA) models bring robots the "brain" for semantic understanding, reasoning, and planning; and the multi-layered data pyramid composed of video priors to simulation synthesis continues to supply steady fuel for the continuous evolution of embodied AI. However, the industry still faces one critical bottleneck: data. Compared to large language models and autonomous driving, embodied AI still faces a massive data

OpenClaw released a new version on March 22, introducing the ClawHub plugin marketplace, support for MiniMax M2.7, GPT-5.4-mini, and GPT-5.4-nano models, and the launch of a single-agent inference feature. It also integrates OpenShell and SSH sandbox environments, along with access to multiple search tools.

Cursor released Composer 2, its third-generation programming model, on March 20th, with pricing significantly reduced to $0.50 per million input tokens and $2.50 per million output tokens, while also launching a faster version. Composer 2 outperforms its predecessor on multiple benchmarks but falls short of GPT-5.4, with improvements primarily stemming from continued pre-training and reinforcement learning of the base model. The model is exclusively available for internal use within Cursor, and its parent company Anysphere has a valuation of $29.3 billion.

OpenAI released new models GPT-5.4 mini and GPT-5.4 nano on March 18, targeting multimodal and programming scenarios with significantly higher pricing. GPT-5.4 mini demonstrates superior performance on coding benchmarks, supports large-scale context and multifunctional operations. Additionally, ChatGPT free users will also experience new reasoning capabilities.

Rakuten Group launched Rakuten AI 3.0, a Japanese-specialized large language model, on March 17, claiming to surpass GPT-4o on multiple Japanese language tests. However, netizens discovered that the model may be based on DeepSeek development and raised questions about its pro-China stance and the authenticity of its independently developed technology, sparking discussions.

Rakuten Group announced the release of Rakuten AI 3.0 on March 17, a high-performance AI model with 671 billion parameters optimized for Japanese. The company claims that it outperforms GPT-4o on multiple benchmark tests. The model is freely open-sourced under the Apache 2.0 license, fine-tuned based on the DeepSeek V3 model, and received training computational support from the Japanese government.

Gate News Announcement: On March 8th, Openclaw recently released version 3.7 beta, mainly adding support for large language models GPT 5.4 and Gemini Flash 3.1.

OpenClaw founder Peter Sternberg shared benchmark results for PinchBench, showing that Gemini 3 Flash ranks first in proxy tasks with a success rate of 95.1%. Domestic models minimax-m2.1 and kimi-k2.5 placed second and third with success rates of 93.6% and 93.4%. Other models such as Claude Sonnet 4.5 and GPT-4o performed lower.