AI news: April

Another month is behind us and with it another batch of news from the world of artificial intelligence. As usual, it's been busy - new models, new architectures, record-breaking contextual windows, and an emphasis on speed and price.

We chose carefully, focusing on what we think will move AI forward the most.

Release Llama 4

Meta has released Llama 4, a new series of open-weight language models that brings two major new features:

• Moving to a MoE (Mixture of Experts) architecture that activates only a small part of the model - specific "experts" - with each query. The result is both higher speed and lower cost.
• Three different models: The fastest Scout, the Maverick with a million context window and the biggest Behemoth, which is still in the training phase.
• The Scout variant allows for a contextual window of up to 10 million tokens, an extreme shift from commonly available models. A very large context window is still a rather theoretical possibility - models are not yet able to "equip" all contextual information at this scale.
• Useful resources:

Release of GPT-4.1 models

OpenAI is coming out with a new iteration of its core model: the GPT-4.1:

• primarily available via API
• Three different models (4.1, faster and weaker Mini and fastest Nano)
• cheaper than GPT-4o, but at the same time a bit slower - the bottleneck is the response speed
• capable of handling up to one million tokens
• significantly better at following instructions

The model works very well with long texts and their context. Together with GPT-4.1 OpenAI introduced a new benchmark for MRCR (Multi-round Co-reference Resolution). The new GPT-4.1 Nano tier is currently the fastest of all, but also the least powerful.

OpenAI launches multimodal models o3 and o4-mini

These are the most advanced reasoning models yet. Both are available for paying users and can be used via API.

• o3 achieves "state of the art" results on really complex benchmarks such as Codeforces or SWE-bench
• o4-mini is a smaller but faster reasoning model
• Both models are specifically trained on tool usage (function calling), suggesting their possible use in intelligent agents.