Tencent Hunyuan releases technical report on Hy-Nomas and launches Hy-Nomas-Bench benchmark for world models

Amap, a division of Alibaba Group focused on mapping and geolocation, has introduced Lambda-0. This model is positioned as the first foundational model for humanoid robots, created based on observed human behavior. The research was conducted in collaboration with Alibaba's ATH business group, and the code is available on GitHub under the Apache 2.0 license alongside a scientific paper on arXiv.

The core idea is that a humanoid can learn walking, posture, and two-handed manipulation simultaneously, even if the robot is absent during data collection. The team asserts that many vision, language, and first-person action-based models focus only on hands and limbs, ignoring crucial information about the torso, body, and posture needed for long-term tasks in open environments. Using teleoperation of real robots to fill this gap is costly and slow to scale.

Instead, Lambda-0 utilizes a whole-body data collection method without the robot's involvement, which Amap calls 'robot-free whole-body data collection.' Specialists wear consumer devices and record movements in everyday settings. For a single demonstration task, collecting one segment took about 29 seconds, which is significantly more efficient compared to the approximately 100 seconds required when using a physical robot, demonstrating a 3.4 times efficiency gain. Through this process, the team created the HumanVerse-500 dataset, containing 500 hours of human behavior data in open environments, synchronized across vision, body movement, and hand actions.

The training process occurs in three stages: hand interaction, full-body coordination, and finally, embodiment in a humanoid body. To account for differences between human and robotic bodies, as well as end-effector errors, the model employs a viewpoint alignment scheme and a coarse-to-fine action alignment method while maintaining full-body coordination and improving contact accuracy.

According to the presented paper, Lambda-0 demonstrated state-of-the-art results in the SIMPLE simulation environment, achieving about 90 percent success rate, as well as in real-world robot tests. In seven long-horizon, full-body mobile manipulation tasks, the model successfully transferred its learning to the Unitree G1 using only human data. The team reports a real-robot success rate of 62.5 percent, which is 22.5 percentage points higher than the strongest baseline. With a constant number of robot demonstrations, increasing the volume of human behavior data generally continued to improve performance on the real robot, which the team presents as evidence of the scalability of humanoid training based on human data. These figures are taken from the team's paper and have not yet been independently reproduced. For Amap, the release of this model complements their recent work on 3D city world models.

Tencent Hunyuan has introduced the second version of its memory framework for AI agents, Hy Memory 2.0, and made all three levels of the stack open source. These levels include a specialized memory model, the framework itself, and a local memory application called MemHub. Users can apply these components either individually or in combination.

Tencent reports that Hy Memory is already being used by internal teams in areas such as finance, news agencies, knowledge management, office agents, social entertainment, and gaming.

The framework is responsible for storing, updating, and developing three types of memory: chat, working, and code. Developers integrate it into their agents via an SDK available on PyPI under the name hy-memory 2.0.2. The new features in version 2.0 are working and code memory. They record the task goal, decisions made, intermediate results, and useful lessons, as well as the conditions applicable to these lessons. Before resuming a long-term task, the system first determines the current task, then retrieves relevant decisions, operational experience, and previous work history.

Chat memory now groups disparate facts into events, indicating time, location, and participants, merges duplicates, and resolves contradictions before passing context to the model.

Framework Testing Results

Tencent stated that it ranked first among open-source frameworks, including Mem0, MemOS, and EverOS, in four different tests. According to the company, working memory showed an average score of 55.31% in trip planning and shopping tasks on the MemoryArena platform. Code memory achieved 64.2% accuracy in question answering and 35.6% task resolution on CodeMemoryBench, as well as 28.3% task resolution on SWE-ContextBench. Chat memory demonstrated 72.6% accuracy on LoCoMo while reducing the number of tokens per question from 1159 to 703, representing a 39.3% decrease.

The layer model, named Hy-Memory-30B-A3B, is a 'mixture of experts' type model with approximately 3 billion active parameters, trained for memory retrieval and conflict resolution. This model is evaluated in a technical report on a fixed sample of 20% of benchmarks across all four memory systems. On this sample, it showed a macro average of 55.49, slightly exceeding the scores of Claude Opus 5 (54.62) and DeepSeek V4.1 Flash (53.64). The model weights are available on Hugging Face.

The MemHub application is a locally oriented tool that includes a command line and a Mac client. After authorizing connected tools, the user imports past sessions into a unified local memory store and supports synchronization of new data. Tencent notes that MemHub is currently integrating with 10 AI tools, including WorkBuddy, Cursor, Codex, Qwen Code, OpenCode, and DeepSeek Harness.

The repository license file on GitHub states that Hy-Memory is released under the Apache 2.0 license.