**A week ago, we talked about the release of Llama 4. Now, the first critical voices are coming from the AI world. And this is justified for several reasons.**

<figure class="">
  <img src="/assets/img/news/2025-04-08-llama-4-does-not-meet-the-expectations-of-the-ai-community/header_img.webp"
       alt="Image with the writing Llama 4"><figcaption>
      Image: Generated with AI

    </figcaption></figure>


## The details

* The models are very large (Llama 4 Scout: 109B total parameters, Llama 4 Maverick: 400B total parameters, Llama 4 Behemoth: 2T total parameters).
* Many AI developers have expected smaller models, including us. Similar to Llama 3. These large models are not suitable for many use cases. Maybe smaller models will come, but Meta doesn’t mention it in the blog article about Llama 4.
* In addition, Meta cheated with the results of Llama 4 in the Chatbot Arena. In the Chatbot Arena, people evaluate the results of the chatbots. This results in the ELO score. According to the Chatbot Arena, Llama 4 Maverick achieved a score of 1417. This score is better than GPT-4o. However, this is a special Llama 4 version and not the one that Meta released.
* Meta released Llama 4 on a Saturday, very unusual. The current model size selection suggests that Meta released the model hastily. In addition, a reasoning model is missing. According to [Yan LeCun](https://x.com/Yanlecunn/status/1908762146957385851){:target="_blank" rel="noopener"}, it comes later.

<figure class="">
  <img src="/assets/img/news/2025-04-08-llama-4-does-not-meet-the-expectations-of-the-ai-community/chatbot_arena.webp"
       alt="Chatbot Arena Results"><figcaption>
      Chatbot Arena Results (Screenshot by authors)

    </figcaption></figure>


**Our thoughts**

We are a fan of the Llama 3 models, especially for local execution. Llama 4 disappointed us because there are no small models. Additionally, according to the [license](https://www.llama.com/llama4/use-policy/){:target="_blank" rel="noopener"}, individuals or companies based in the EU are not allowed to use the multimodal features of Llama 4.

It’s also interesting to know that Meta’s AI chief, Joelle Pineau, is leaving Meta after eight years. Reportedly, there were internal issues with the training of Llama 4.

**More information:** 🔗 [Meta AI](https://ai.meta.com/blog/llama-4-multimodal-intelligence/){:target="_blank" rel="noopener"} \|  [Heise Online](https://www.heise.de/en/news/Meta-cheats-on-Llama-4-benchmark-10344087.html){:target="_blank" rel="noopener"} \| [AP News](https://apnews.com/article/meta-ai-research-chief-stepping-down-joelle-pineau-c596df5f0d567268c4acd6f41944b5db){:target="_blank" rel="noopener"}

<div class="ad-banner" style="margin-bottom: 0.7rem;">
    <hr class="hr-text" data-content="Advertisement*">
    <a href="https://tinztwinshub.com/out/reclaimai/" target="_blank" rel="sponsored nofollow noopener"><img src="../../assets/img/ads/reclaimai.webp" alt="Reclaimai Partner*" nopin="nopin"></a>
    <small style="display: block; margin-bottom: 0.5rem; margin-top: 0.5rem;"><strong>✨ Read without banner ads? </strong><a href="https://steady.page/en/magic-ai/about" target="_blank" rel="noopener">Become a member</a> or <a href="https://steady.page/en/log_in?publication=magic-ai" target="_blank" rel="noopener">log in</a></small>
</div>

## Magic AI tool of the week

The popular language learning app Duolingo offers video calls with an AI avatar. So, you can talk with an AI avatar in real-time. Duolingo uses the GPT-4o model from OpenAI for this functionality.

Take a look at the following demo video 👇🏽








<!-- Courtesy of embedresponsively.com //-->

  <div class="responsive-video-container">
    <iframe data-name="youtube" data-src="https://www.youtube-nocookie.com/embed/vBjfn2Pb6DI" frameborder="0" webkitAllowFullScreen mozallowfullscreen allowfullscreen></iframe>
  </div>



AI is making its way into the education sector and creating real added value. We use the Duolingo app every day to improve our English skills. The new AI feature is available as a paid subscription.

👉🏽 [Learn more about Duolingo.](https://www.duolingo.com/){:target="_blank" rel="noopener"}

## Hand-picked articles of the week

* [Build a Local AI Agent to Chat with Financial Charts Using Agno](https://tinztwinshub.com/investment-research/build-a-local-ai-agent-to-chat-with-financial-charts-using-agno/)
* [Build a Local RAG App to Chat with Earnings Reports](https://tinztwinshub.com/investment-research/build-a-local-rag-app-to-chat-with-earnings-reports/)
* [Build an OpenAI-Compatible API for Open-Source LLMs from Scratch](https://tinztwinshub.com/software-engineering/build-an-openai-compatible-api-for-open-source-llms-from-scratch/)

<div class="ad-banner">
    <hr>
    😀 Do you enjoy our content? If so, why not support us with a small financial contribution? As a supporter, you can comment on newsletter editions (e-mail version) and read our website without banner ads.

    <div style="text-align: center; margin-top: 0.7rem;">
        <a href="https://steady.page/en/magic-ai/about" class="btn btn--primary" target="_blank" rel="noopener">Support us</a>      
    </div>
</div>

<div class="ad-banner">
    <hr class="hr-text" data-content="Our Merch Shop">
    <a href="https://shop.tinztwinshub.com/merch/"><img src="../../assets/img/ads/ai_and_coding_merch.webp" alt="AI and Coding Merch" nopin="nopin"></a>
    <small style="display: block; margin-bottom: 0.5rem; margin-top: 0.5rem;"><strong>✨ Read without banner ads? </strong><a href="https://steady.page/en/magic-ai/about" target="_blank" rel="noopener">Become a member</a> or <a href="https://steady.page/en/log_in?publication=magic-ai" target="_blank" rel="noopener">log in</a></small>
</div>