# Eleven of nineteen open language models use Apache 2.0

> The rest split into MIT, modified MIT and four custom licences. Size does not predict which, so read the licence file.

Published 2026-10-23

Canonical: https://shipped.anth.us/stories/open-weights-licences/

Shipped's open-weights section lists nineteen language models. Eleven ship under Apache 2.0. That count covers this list, not the whole field.

The Apache group spans every size. Beam has 501B parameters [1]. Kolibri 1 has 78B [2]. Yandex's base model has 80B [3]. Qwen3.8-27B has 27B [4]. Mellum2.1 has 12B [5]. MiniCPM5-2B has 2B, and its training data is open [6]. Granite 4.1 8B [7], Ministral 3 [8], Step 3.7 Flash [9] and Olmo 3 [10] join them. Gemma 4 joins too. Its card shows Apache 2.0, and the item notes that earlier Gemma terms were stricter [11].

Two use plain MIT. DeepSeek V4.1-Flash has a 552B backbone [12]. MiMo-V2.6-Pro-RL has 1.02T parameters [13].

Two use a modified MIT. Mistral Medium 3.5 is open for commercial and non-commercial use, with exceptions for companies with large revenue [14]. Kimi K2 Thinking has both code and weights under a Modified MIT License [15].

Four use custom licences. Kimi K3 has its own Kimi K3 License [16]. GLM-5.3 uses a licence named glm-5.3 [17]. Liquid AI's LFM2.5-24B-A2B uses lfm1.0 [18]. NVIDIA's Nemotron 3 Nano uses the NVIDIA Nemotron Open Model License [19].

Size does not predict the licence. Kimi K3, at 2.8T, is the largest on the list and is custom. MiMo, at 1.02T, is MIT. Beam, at 501B, is Apache. LFM2.5, at 24B, is custom.

So a licence name is only a start. The custom group holds both the largest model on the list and one of the smaller ones. Open weights can still carry a revenue limit or a bespoke text. Read the file before you build.