New Multimodal Model Emerges as Kimi Accelerates Vision Understanding Push After Funding

A Mysterious Model Appears, Stirring the Arena Again

Recently, an unknown model named “Kiwi-do” appeared on an anonymous large model arena, quickly attracting widespread attention from the developer community. A user asked the model about its identity on social media, and it described itself as coming from a well-known AI startup team, with training data up to January 2025. This detail suggests the model is not a simple variant of an existing public version, but possibly a newer-generation model with more recent training and more complete capability iterations.

Meanwhile, the team completed a funding round of approximately $500 million (about 3.5 billion RMB) in late 2025, with participation from multiple leading institutions and additional subscriptions from existing shareholders, reaching a post-money valuation of $4.3 billion. An internal letter mentioned that after this round, the team’s cash reserves reached 10 billion RMB, with funds mainly used to “aggressively expand GPU resources” to accelerate training and R&D of the next-generation foundation model. Given this timing, the unknown model suddenly appearing on the arena is very likely one of the first technical outputs after the new funding landed.

Community Testing Clues Point to a Multimodal Version

The blogger who first discovered the model compared Kiwi-do with the team’s already-launched K2-Thinking model, focusing on SVG drawing capability. The tests included two drawing tasks: “a pelican riding a bicycle” and “a game controller.” From the results, Kiwi-do’s SVG output differed significantly from K2-Thinking, indicating they are not different session states of the same model, but have substantive differences in parameters or training objectives.

The blogger then further evaluated Kiwi-do using the VPCT (Visual Physical Comprehension Test) benchmark. VPCT requires models to reason based on understanding image content combined with physical laws, such as determining which container a ball will eventually fall into in a specific device. Such tasks not only test visual encoding capability but also require spatial relationship modeling and physical intuition reasoning, placing high demands on the overall level of multimodal models. Test results showed that Kiwi-do correctly answered all questions.

This result sparked community discussion. Some netizens speculated that Kiwi-do is likely the multimodal model K2-VL mentioned by the team in a previous AMA event. Others suggested it might be a smaller-parameter experimental model, but its performance on visual physical reasoning already has reference value. Other domestic media reported that the team plans to launch a new multimodal model in Q1 2025, possibly named K2.1 or K2.5, which may point to different internal codenames for the same model as K2-VL.

Technical Challenges and Evaluation Dimensions of Multimodal Understanding

Regardless of Kiwi-do’s final identity, the evaluations and discussions around it point to a core question: where exactly are the capability boundaries of vision-language models (VLMs)? From a technical perspective, current multimodal model development mainly faces challenges at several levels.

The first is the fusion of visual encoding and language alignment. Early VLMs mostly adopted a simple concatenation approach of “frozen visual encoder + language model,” with limited alignment precision between visual features and text semantics. Subsequent research found that introducing trainable visual adapters, cross-modal attention modules, and large-scale interleaved image-text data can significantly improve alignment quality. However, different architectures vary greatly in their preservation of fine-grained visual information, especially when handling high-resolution images, complex scenes, and overlapping objects, where capability fluctuations are pronounced.

The second is spatial reasoning and physical commonsense modeling. The rise of benchmarks like VPCT is precisely to fill the gap in “deep reasoning” left by traditional VQA (Visual Question Answering) tasks. Traditional VQA mostly stays at the level of recognition and description, such as determining object color, quantity, or scene category; while physical understanding tasks require models to predict event outcomes, which demands that models build an “internal physics simulator” capability on top of visual perception—even without a real physics engine, they must be able to approximately deduce causal chains in semantic space.

The third is the quality and distribution of training data. Reasoning capability of multimodal models depends heavily on the diversity and difficulty of training data. If the training set contains a large number of simple “image captioning” samples, the model’s capability on complex visual reasoning will be hard to fully stimulate. Conversely, if synthetic data with spatial annotations, trajectory information, and physics simulation environments is introduced, it can effectively improve the model’s physical intuition, but generating such data is costly, and balancing data scale with quality is an ongoing optimization problem.

The fourth is the improvement of evaluation systems. Currently, industry evaluation of multimodal models is still in a development stage. Beyond VPCT, there are benchmark sets focusing on different aspects such as chart understanding, document parsing, UI navigation, and embodied intelligence. A high score on a single benchmark does not equate to the model being usable across all scenarios; it must be combined with multi-dimensional evaluation, human feedback, and real task validation to give a relatively objective conclusion about model capability.

Application Prospects Driven by Multimodal Models

If Kiwi-do is indeed K2-VL or an early version of it, then its passing the VPCT test is significant not only for academic metrics but also for opening more imaginative space for practical applications. Multimodal agents are one of the most promising directions for deployment. Traditional text agents can only handle structured or semi-structured data, while multimodal agents can directly understand screenshots, charts, process diagrams, camera feeds, and other information, thereby executing more complex operational tasks.

Taking automated testing as an example, a model with visual physical understanding capability can infer the expected responses of interaction behaviors such as clicking, dragging, and scrolling while looking at the UI interface. Even when facing closed systems without DOM trees or API interfaces, it can complete a certain degree of automated operations. Another typical scenario is intelligent document processing. For documents containing complex tables, flowcharts, formulas, and handwritten annotations, if the model can simultaneously parse visual layout and semantic logic, it will greatly improve the accuracy and robustness of information extraction.

Additionally, in fields requiring spatial perception and causal reasoning such as robot manipulation, autonomous driving assistance, and educational tutoring, multimodal models with VPCT-level capability also have high transfer value. Although these scenarios still face obstacles such as long-tail problems, safety constraints, and real-time requirements, once the model’s underlying capability is established, the adaptation cost for upper-layer applications will gradually decrease.

Bidirectional Drive Between Funding Rhythm and Technical Evolution

From an industry-wide perspective, competition in multimodal models is shifting from a “parameter race” to a “capability race.” The era of simply competing on parameter count or training data scale is gradually passing, and how to balance visual understanding, logical reasoning, tool calling, and long-context modeling has become the core contradiction in current R&D. Computing power expansion brought by funding is certainly important, but more critical is how to design effective training paradigms and data strategies under abundant computing conditions.

Another direction worth watching is model miniaturization and efficient deployment. Kimi has stated that new models may be anonymously tested on the arena in versions with different parameter scales. This strategy helps collect diverse user feedback before official release while avoiding unnecessary impact on model reputation from single-point results. From a technical perspective, compressing model size through distillation, quantization, pruning, and other methods while maintaining core capabilities can make multimodal capabilities more easily embedded into edge devices and on-device applications, further expanding commercialization scenarios.

Conclusion: Long-Term Considerations Beyond the Arena

The emergence of Kiwi-do is a microcosm of the resonance between funding and technical iteration. It reflects leading AI teams’ determination to accelerate model iteration with new rounds of computing power support, and also reveals that the evolution path of multimodal models from “can see” to “can think clearly” is not yet complete. For developers, following model dynamics in anonymous arenas is interesting, but more important is establishing rational judgment about multimodal capabilities—beyond evaluation data, considering model stability, controllability, and maintainability in real business scenarios is the true deciding factor in technology selection.

In the coming quarters, the release pace of multimodal foundation models may further accelerate, and evaluation standards around visual physical reasoning, complex document understanding, and multi-step agent tasks will continue to improve. By then, the market will no longer focus solely on “which model scores higher,” but on “which capability system can truly integrate into products and solve users’ actual problems.” And all of this begins with those mysterious models with seemingly arbitrary names in the arena.