April

3 events

Editors' summary

Reasoning models pick their own tools and think with images

On 16 April, OpenAI released the reasoning models o3 and o4-mini. They combine tools — web search, running Python, analysing images — on their own initiative, and pull images into the reasoning itself, zooming and rotating as they think.

Meta released Llama 4 on the 5th. Questions about whether the version submitted for evaluation matched the one shipped left it poorly received.

Google added Gemini 2.5 Flash. Each lab was filling out its range of reasoning models.

On the 29th, Alibaba released Qwen3 with dense models from 0.6B to 32B alongside two mixture-of-experts models, adding a second Chinese line to the labs publishing weights.

This block is written by the editors. It is kept separate from the sourced record below.

Record3 events

5
ModelMeta
Llama 4 moves to MoE, and runs into a benchmark row
Meta released Llama 4 Scout and Maverick, moving from the dense Transformers of the previous generation to a sparse mixture of experts: Scout at 109 billion parameters with 17 billion active, Maverick at 400 billion with the same 17 billion active. Both took images as well as text, and Scout claimed a then-unprecedented 10-million-token context window. A two-trillion-parameter Behemoth was described as still in training and was not released. The launch landed on a Saturday, ahead of Meta's own LlamaCon later that month, which struck observers as odd, and it was followed by a row over evaluation when the version that scored well on LMArena turned out not to be the same as the weights that shipped. For the company that had carried the open-weight banner, it read as a stumble at the moment the field was catching up.
16
ModelOpenAI
OpenAI releases reasoning models o3 and o4-mini
OpenAI released OpenAI o3 and o4-mini, the latest models in its o-series of reasoning models. o3 was positioned as the company's most powerful reasoning model, while o4-mini offered strong performance at lower cost and latency. They were the first reasoning models able to agentically use and combine every tool within ChatGPT, including web search, Python, image analysis, and image generation, and introduced the ability to "think with images" by integrating them into the reasoning process.
29
ModelAlibaba
Alibaba releases Qwen3 across dense and MoE sizes
Alibaba released Qwen3 with open weights: six dense models from 0.6B to 32B alongside two mixture-of-experts models (30B total with 3B active, and 235B with 22B active), trained on 36 trillion tokens — twice Qwen2.5. The models could switch between thinking and non-thinking modes, with gains claimed in reasoning, instruction following, tool use, and multilingual work. Three months after DeepSeek drew the world's attention, a second Chinese line arrived with a full range of sizes. Qwen went on to become the largest open-weight foundation by derivative models and downloads; in August 2026 it was reported to have passed 3 billion cumulative downloads on Hugging Face, ahead of both Meta and Google.

↑ 2025

Llama 4 moves to MoE, and runs into a benchmark row

OpenAI releases reasoning models o3 and o4-mini

Alibaba releases Qwen3 across dense and MoE sizes