更新日時：　投稿日時：2023-10-27

Today, we at Cognition Labs are thrilled to introduce Nexus-3, our latest and most powerful AI model. Nexus-3 represents a significant leap forward in our mission to build safe and beneficial artificial intelligence. It was designed from the ground up to excel at complex reasoning, operate with groundbreaking efficiency, and seamlessly understand a combination of text, images, and audio.

This release isn't just an incremental update; it's a new benchmark for what's possible.

What's New in Nexus-3?

We focused on three core pillars during the development of Nexus-3: advanced reasoning, computational efficiency, and deeper multimodal capabilities.

Advanced Reasoning and Problem-Solving

Nexus-3 demonstrates a dramatically improved ability to understand nuance and solve multi-step problems that require logical deduction. Where previous models might provide surface-level answers, Nexus-3 can "think" through a problem, breaking it down into smaller parts and synthesizing a coherent solution.

For example, you can present it with a complex business challenge, complete with market data and constraints, and ask it to draft a strategic plan. It can identify potential risks, propose mitigation strategies, and justify its recommendations with clear, logical arguments.

Unprecedented Efficiency

State-of-the-art performance has historically come at a high computational cost. With Nexus-3, we've broken that paradigm. Thanks to a novel model architecture and optimization techniques, Nexus-3 delivers its superior performance while using 40% less compute than our previous leading model, Nexus-2.

This efficiency makes Nexus-3 faster, more cost-effective for our partners to use, and opens the door for powerful AI applications to run on a wider range of devices.

Deep Multimodal Understanding

The world isn't just text, and a truly helpful AI needs to understand that. Nexus-3 has native, deeply integrated multimodal capabilities. It doesn't just process different types of inputs; it comprehends the relationship between them.

• Analyze a user interface mockup image and generate the corresponding code.
• Listen to a recorded meeting and produce a structured summary with action items assigned to the correct people.
• Look at a complex scientific diagram and explain the process it illustrates in simple terms.

This opens up a new class of applications that feel more intuitive and integrated into our daily workflows.

Setting a New Standard: Performance Benchmarks

We've tested Nexus-3 against a wide range of industry-standard academic benchmarks. It sets a new state-of-the-art record on the majority of them, showcasing its robust capabilities across different domains.

Here’s a brief look at how it compares to other leading models on key benchmarks for reasoning and knowledge:

| Model | MMLU (General Knowledge) | GSM8K (Math Reasoning) | HumanEval (Coding) | |---|---|---|---| | Leading Competitor Model | 86.4% | 92.0% | 74.4% | | Nexus-2 (Our Previous SOTA) | 87.1% | 93.5% | 78.1% | | Nexus-3 (New) | 90.2% | 97.1% | 88.4% |

These results underscore the significant advancements in reasoning and problem-solving baked into the core of Nexus-3.

How to Access Nexus-3

We believe in getting powerful tools into the hands of builders as quickly as possible. Starting today, Nexus-3 is available in two ways:

1. For Developers: Nexus-3 is now the default model powering our API. Developers can get started immediately by visiting our Developer Portal. The improved efficiency means API calls are faster and more affordable.
2. In Our Products: Nexus-3 is being rolled out to our flagship application, Cognition Assistant, for all users starting this week. You'll notice more accurate, faster, and more helpful responses.

The Road Ahead

The launch of Nexus-3 is a major milestone, but it's just one step on a long journey. We are more committed than ever to building AI that is not only powerful but also safe, aligned with human values, and accessible to everyone.

We can't wait to see what you build with it.

---

The Cognition Labs Team