Open-weight models

Worked check: Mistral Large 4 is about 1T parameters with ~49B active under Mistral’s license — the downloadable file is a separate fact from any one leaderboard row.

Open weights are not one license. Apache 2.0, as with Beam and EmbeddingGemma 2, is a license many companies already know how to read. Meta's Muse license, as with Muse Spark Open, is a different document. Read it before treating the file as something a product can ship.

Size is the other gate. On October 29, Mistral released Large 4 (Le Chonk): about 1 trillion parameters, roughly 49 billion active, a 1 million token context, with Mistral citing 41 percent on Terminal-Bench 4.0. Spark Open is 816 billion parameters with 42 billion active, and Meta's own setup is two M5 Ultra Mac Studios. Beam is 501 billion with 23 billion active, under Apache 2.0. MiniMax-M3.1-Flash lists about 428 billion parameters with 23 billion active, and the license name on the card is MiniMax Community. EmbeddingGemma 2 is 740 million and is meant to run on the device that holds the files. The word open does not say which of those machines you have.

In this briefing