Tencent Yuanbao Application for HarmonyOS Released with Hunyuan Hy4 Preview Version

TaichuAI has made the ZDTaichu5.0-9B model publicly available. This multimodal foundation model has approximately 9 billion parameters and is designed for general visual understanding, spatial reasoning, agent tool usage, and embodied AI research.

The model's architecture combines the Qwen3.5-9B language backbone with the C-RADIOv4-H vision encoder. The model accepts text input, as well as one or more images and videos of any resolution, supporting a context length of up to 128 thousand tokens. The release, summarized by TMTPost on September 15th, focused on real-world spatial perception, transformations between different views, and task planning for embodied systems.

According to the public model card, TaichuAI demonstrates state-of-the-art results among approximately 10-billion general VLMs across a wide range of visual tasks, while also expanding its capabilities in spatial, embodied, and agentic scenarios. Noteworthy metrics in space and embodiment include ViewSpatial 62.50, MMSI-Bench 47.20, MindCube-tiny 78.27, ERQA 48.00, and RoboSpatial 56.00.

Regarding agent and instruction sets listed in the evaluation card notes, TAU2-Bench achieves a score of 87.70, and the average Claw-Eval score is 71.40, while IFEval shows 93.70. The entropy-prioritized adaptive recursive reasoning mechanism is described as a mechanism that allocates additional refinement steps for latent variables for more complex tokens.

The model's capabilities cover Optical Character Recognition (OCR) and document understanding, math on images, fine-grained 2D relationships, multi-view association, 3D scene and perspective perception, as well as tracking multiple images and videos within a long context window, including multi-step tool use planning. Actual tool execution remains the responsibility of the host application.

The spatial learning topics listed on the card include relative relationships, dense counting and framing, camera motion and depth ordering, egocentric versus allocentric views, and high-level capability definition and action planning for VLA-style adaptation.

The model weights are hosted on Hugging Face at TaichuAI/ZDTaichu5.0-9B under the NVIDIA Open Model License, while retaining the Apache-2.0 notices from Qwen3.5. A custom branch of vLLM 0.26.0 and a Docker image from TaichuAI are used for serving to provide OpenAI-compatible endpoints, with recommended sampling settings for spatial grounding and general tasks.

As with other vendor-released models, external laboratories should view benchmark scores as reported with the specified prompts and judges until independent reproduction results emerge. However, the combination of a mid-sized open multimodal checkpoint, a focus on space and agents, and ready vLLM packaging provides researchers with a concrete artifact for evaluation.

Shanghai-based startup EBKernel unveiled Cog-WM 1.0 at the embodied brain intelligence session during the Puzhang Innovation Forum in 2026. The company positions this release as a cognitive world model built on latent space prediction rather than pixel reproduction.

According to the company's statement, this architecture borrows organizational concepts from human cognitive maps, such as selective memory updating, goal-conditioned encoding, and multi-horizon prediction. These ideas are combined with a JEPA-style joint embedding objective, enabling robots to plan using abstract spatial and state features instead of reconstructing raw frames.

In terms of navigation, the Cog-WM Nav 1.0 model was tested without using a pre-built map. On a subset of HM3D-ObjectNav, EBKernel reported an increase in success rate from 78.50% to 86.89% compared to the BSC-Nav baseline published in Nature Communications. This represents an absolute gain of 8.39 points (approximately 10.7% relative), and the SPL metric increased from 47.70 to 48.35.

The navigation component maintains explicit spatial memory and predicts subgoals for exploration in the latent space. It then balances between goal semantics and path cost, which is particularly useful when the target is outside the current field of view and the robot needs to decide where to look next.

The manipulation branch, Cog-WM Manip 1.0, trains multi-scale state and value-modulated experience prediction, providing a link between short-horizon action effects and long-term task progress. Using a unified replay protocol, EBKernel claims that the model outperforms massively pre-trained baseline models like pi0.5 by approximately 16% across three main manipulation datasets, including more complex RoboTwin 2.0 settings.

Analysis of component influence on LIBERO-Plus showed an additive effect: the policy alone achieved nearly 80%, local prediction reached 81.6%, multi-horizon prediction reached 82.0%, and the full system reached 84.6%.

The company states deployment capabilities cover wheeled and quadrupedal humanoids for mapless navigation, route planning, spatio-temporal memory retrieval, spatial question answering (QA), and object searching. Manipulation was tested on wheeled humanoid bodies, while quadrupedal navigation was noted on client objects for inspection or patrolling.

EBKernel, founded in mid-2025, promotes a product thesis centered on lifelong learning with low data requirements and high generalization ability. However, independent outdoor durability metrics and long-horizon field results remain within the confines of the company's internal benchmarks, so laboratories are advised to consider the published SR metrics and manipulation changes as the primary comparison set until third-party work emerges.