Hey {{first_name}},

2025 is the year of AI agents. And as they are powered by LLMs, the better LLMs do the better are the agents. It is about their performance, but also features like context window length, latency, and hallucination rates.

Today on the docket:

The Chinese company Alibaba has released a reasoning LLM with the slightly cryptic name QwQ-32B. The latter part stands for its number of parameters: 32 billion. For comparison DeepSeek-R1 has 671 billion (20x).

This is matters a big deal, because the more parameters the higher compute costs and latency.

This relatively small model outperforms very large models significantly in math (AIME benchmark). It is ahead of R1, and OAI’s o1.

In other benchmarks it performs great as well.

It is open source: download it and work with it (65GB) here:

With QwQ, we see another 90% price cut in word generation.

Llama 4 was set to be released in January but R1 made them to rethink/ rework their model.

With Llama 4 they have to release something stronger than QwQ-32b. That’s a must. But in the meantime DeepSeek announced that R2 is coming in May 2025. 🙈

Turbulent times ahead, but all good for the consumer side of it.

It’ll be cheaper and more performant.

When working with tools like Cursor AI, Windsurf AI, or Anthropic you want to connect it reliably to various data. MCP is a protocol to streamline that process.

You no longer have to struggle with setting up MCP servers or worry about server reliability, availability, and authentication complexity.

Composio has built the largest source of fully managed MCP servers, with complete auth support.

What they have to offer

That is a big deal, as it saves a lot of time.