According to CoinMeta, the luxurious founding team of Recursive announced the release of their first model, Flint. This model is specifically designed for local programming on agent and has a total of about 40 billion parameters, with only 2.5 billion activated at a time. The team developed this model from scratch in less than 4 months. In May of this year, Recursive received $650 million in financing, raising its valuation to $4.65 billion. The team's eight co-founders all have significant backgrounds, including the first author of Vision Transformer, Alexey Dosovitskiy; former research director of Meta Fair, Tian Yuandong; former head of Google DeepMind, Tim Rockt; and schel. Tests published by Recursive show that on a M4 Pro MacBook Pro with 48GB of memory, Flint scores about 81% in comprehensive programming and instruction evaluation, slightly lower than Qwen3.8's 85%, but its long-context comprehensive generation speed is nearly 7 times that of the latter. The model weights have been made open according to the Apache 2.0 protocol.