FUMO Lab Officially Launches Fusion Model, Setting New Benchmark for Frontier Model Capabilities | Yunqi Capital

As tasks grow more complex, an Agent is increasingly like a toolbox: not every decision about when to reason, when to write code, or when to invoke tools needs to be handed to the same model.

To address this, FUMO Lab has officially launched two fusion model products — Fusion Max and Fusion Code — targeting general complex tasks and software engineering / code Agent scenarios respectively.

FUMO Lab is a Singapore-based frontier model lab and the sister company of Xiaosu Technology, a Yunqi Capital portfolio company. What Fusion aims to do is enable different models' capabilities to be combined more effectively within a single task, based on task progress and current state.

---

The following is republished from "Xiaosu Technology"

On four top public benchmarks — reasoning, industry, programming, and science — Fusion Max and Fusion Code both claimed first place, surpassing GPT 5.6, Fable 5, and Moonshot AI K3 respectively.**

Today, Singapore-based frontier model lab FUMO Lab — also sister company to Xiaosu Technology — officially launched two fusion model products, Fusion Max and Fusion Code, opening them for invite-only beta access. According to FUMO Lab's white paper, the Fusion products surpassed existing top models on public benchmarks including GPQA Diamond, τ³-Banking, and Terminal-Bench v2.1, while sitting on the performance-cost Pareto frontier, reducing full-task costs by approximately 30%–40%.

When Agents handle long-horizon tasks, the most suitable model capability shifts continuously as new evidence, tool outputs, and task states emerge. Fusion maintains a shared state, constantly assessing what's still missing, which capability to invoke next, and whether further computation is worthwhile — ultimately delivering consistent results through a unified interface. Of the two products, Fusion Max targets general complex tasks; Fusion Code targets software engineering / code Agent scenarios.

From Yunqi Capital's perspective, Fusion's core value lies in a compounding system capability: every improvement at the foundation model layer directly raises Fusion's ceiling; as the system's understanding of task structures and model capability boundaries deepens, allocation becomes more precise, generating greater value from the same set of models. Better task outcomes and evaluation signals in turn refine the allocation layer, creating a continuously evolving feedback loop.

The important shift is not from "one model" to "many models," but from statically selecting models to adaptively allocating intelligence based on task state.

Fusion Max and Fusion Code are now open for invite-only beta access. To apply for access or read the white paper:

【1. Official website: fumolab.ai ; 2. Long-press to scan QR code】

Fusion Model by FUMO Lab — One fusion, the strengths of many.

融合模型，尽揽众长。