觉得大模型消耗的算力过大，英伟达推出的8B模型Orchestrator化身「拼好模」，通过组合工具降本增效，使用30%的预算，在HLE上拿下37.1%的成 绩。 最近，NVIDIA Research发现，只要经过适当微调，小模型已足以「指挥」大模型 英伟达研究团队的新模型Orchestrator仅有 80 亿参数（8B）的模型，不仅比以往的工具使用类AI智能体准确率更高、成本更低，还能在工具选择上精准对 齐用户的偏好。 在HLE基准测试中，Orchestrator斩获了37.1%的高分，一举超越了GPT-5（35.1%），同时在效率上提升了2.5倍。 在tau2-Bench和FRAMES测试中，Orchestrator同样以大幅优势领先 GPT-5，而其成本仅为后者的30%左右。 在多项指标上，Orchestrator均实现了性能与成本的最佳平衡，并能出色地泛化至未曾见过的工具中。 预印本链接：https://arxiv.org/abs/2511.21689 为什么「强模型+工具」还是不够好？ 面对Humanity's Last Exam（HLE）这类超难综合推理考试，现在的大模型虽然「什么都懂一点」，但一到 ...

Financial Performance - Underlying Revenue from customers reached $256.8 million, a 3.1% year-over-year increase[5] - Recurring Revenue accounted for $154.3 million of the Underlying Revenue[5] - Underlying Cash EBITDA was $43.8 million, resulting in a 17% Cash EBITDA margin[5] - The company ended the year with $58.7 million in cash and no debt[5] - A final FY25 dividend of $13.1 million (2.92c per share) and a special dividend of $8.0 million (1.79c per share) were distributed[5] Segment Performance - EMEA revenue was $186 million in FY25[12] - APAC revenue was $73 million in FY25[12] Future Outlook - The company expects FY26 revenues to be in line with FY25 and Cash EBITDA to be above $50 million[31] - One customer exit will impact FY26 revenue by A$10 million[37] Capital Allocation - A capital return of $73.2 million (16.3 cents per share) was paid in January 2025[28] - A special dividend of $40.0 million (8.92 cents per share) and 1H25 Dividend of $7.2 million (1.6 cents per share) was paid in April 2025[28]