1.1M and 11,965,760 weights (dense Llama architecture)
4,096 tokens
Apache 2.0
Second-generation release of an intentionally tiny family of open-weight bee language models designed to buzz, express moods, perform actions, remember how they are treated within the conversation, and pollinate flowers. BeeLLM 2.0 contains two dense Llama-architecture models distributed directly as F32 GGUF files for llama.cpp: BeeLLM-2.0-1M-Mood, with approximately 1.1 million weights across four transformer layers and a 1,024-token vocabulary, and BeeLLM-2.0-12M-Mood, with exactly 11,965,760 weights across eight layers and a 2,048-token vocabulary. Both use a 4,096-token context window and a custom chat template with <|user|> and <|bee|> roles. The 1M checkpoint occupies approximately 4.49 MB and the 12M checkpoint approximately 47.91 MB. Both were retrained on approximately 3,000 new conversations relative to the first-generation Mood models. Prism-like compression is not involved: these are genuinely very small, full-F32 language models. Responses begin with a mood-and-intensity tag such as [happy:high], [angry:low], or [confused:high], followed by buzzes, short words, and physical actions. The models can react differently to flowers, affection, swatting, bug spray, confinement, apologies, and attempts at immediate lovebombing. The 12M lineage knows that her name is Scintilla, can develop grudges against objects and actions, may file formal complaints against jars, and very rarely stings the user and dies before potentially reviving several turns later. BeeLLM 2.0 gives both variants a richer voice and a favorite flower; the 1M version is friendlier toward gentle messages, while the 12M version is additionally described as kinder to other bots and particularly fond of orange flowers. The broader repository also contains the original 30P and 42P tiny-buzz models, Chaos variants with shouting and stinging vocabulary, and first-generation 1M and 12M Mood checkpoints. All models run anywhere supported by llama.cpp and are small enough to embed in web applications. The official interactive browser demonstration presents the model as a bee with feelings that remembers how the user treats her. No conventional benchmark suite or technical paper is provided. The model card lists its language as buzz, its intended uses as buzzing and pollinating flowers, and its primary limitation as the possibility that it may sting.