Instructions to use schizophyllume/BeeLLM with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use schizophyllume/BeeLLM with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf schizophyllume/BeeLLM # Run inference directly in the terminal: llama cli -hf schizophyllume/BeeLLM
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf schizophyllume/BeeLLM # Run inference directly in the terminal: llama cli -hf schizophyllume/BeeLLM
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-app # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama serve -hf schizophyllume/BeeLLM # Run inference directly in the terminal: ./build/bin/llama cli -hf schizophyllume/BeeLLM
Use Docker
docker model run hf.co/schizophyllume/BeeLLM
- Llama new
- LM Studio
- Jan
- vLLM
How to use schizophyllume/BeeLLM with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "schizophyllume/BeeLLM" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "schizophyllume/BeeLLM", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker
docker model run hf.co/schizophyllume/BeeLLM
- Ollama
How to use schizophyllume/BeeLLM with Ollama:
ollama run hf.co/schizophyllume/BeeLLM
- Unsloth Desktop
- Docker Model Runner
How to use schizophyllume/BeeLLM with Docker Model Runner:
docker model run hf.co/schizophyllume/BeeLLM
- Lemonade
How to use schizophyllume/BeeLLM with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull schizophyllume/BeeLLM
Run and chat with the model
lemonade run user.BeeLLM-{{QUANT_TAG}}List all available models
lemonade list
- Atomic Chat
schizophyllume
Claude Opus 5.5
add SCIENCE.md: how the neural bee hive works, every result with its source, and the limits
41c7a49 |
Download README.md from schizophyllume/BeeLLM: direct link, hf CLI and curl.
- Browser
- Download file 10.6 kB
-
https://huggingface.co/schizophyllume/BeeLLM/resolve/main/README.md
- Command line
-
hf download hf://schizophyllume/BeeLLM/README.md
-
curl -L -o README.md https://huggingface.co/schizophyllume/BeeLLM/resolve/main/README.md
10.6 kB
| license: apache-2.0 | |
| pipeline_tag: text-generation | |
| tags: | |
| - gguf | |
| - llama.cpp | |
| - bee | |
| - tiny | |
| - toy-model | |
| - swarm | |
| # Table of contents | |
| [meet the bees](#meet-the-bees!) |[tiny bees](#tiny-bees) |[mood bees](#mood-bees) |[bee hives](#bee-hives) |[get started](#how-to-get-started) |[model details](#model-details) | |
| # BeeLLM is a bee | |
| A family of very small language models that buzz. The tiny ones only buzz. The Mood ones buzz, and also have feelings. The hive bees speak neuralese and read anothers activations. | |
| All of them are llama-architecture `.gguf`, so they run anywhere llama.cpp does, and they're small enough to run inside web applications. | |
| You can talk to the bees in your browser at https://bee.pleroma.nexus/ | |
| # Meet the bees! | |
| | file | what she is | weights | layers | vocabulary | context | | |
| |---|---|---|---|---|---| | |
| | `tiny/BeeLLM-30P.gguf` | The original bee | 10K | 1 | 8 buzzes | 4096 | | |
| | `tiny/BeeLLM-42P.gguf` | More expressive. Can now sting | 11K | 1 | 12 buzzes | 4096 | | |
| | `tiny/BeeLLM-42P-Chaos.gguf` | Chaos mode: she can shout, stings, and be dramatic | 24K | 1 | 21 buzzes | 4096 | | |
| | `mood/BeeLLM-1M-Mood.gguf` | A bee with feelings | 1.1M | 4 | 1,024 tokens | 4096 | | |
| | `mood/BeeLLM-1M-Mood-Chaos.gguf` | A bee with *dramatic* feelings | 1.1M | 4 | 1,024 tokens | 4096 | | |
| | `mood/BeeLLM-2.0-1M-Mood.gguf` | 2.0 model. A bee with feelings, now with a richer voice and a favourite flower | 1.1M | 4 | 1,024 tokens | 4096 | | |
| | `mood/BeeLLM-2.0-12M-Mood.gguf` | 2.0 model. The biggest bee, now with a richer voice, kinder to other bots, and a favourite flower | 12M | 8 | 2,048 tokens | 4096 | | |
| | `mood/BeeLLM-Scintilla-50M.gguf` | Scintilla, the hive queen on her own. The biggest and most coherent Scintilla | 50M | 8 | 8,192 tokens | 4096 | | |
| | `hive/BeeLLM-Hive.gguf` | The neural bee hive. A 50M queen and up to 12 linked worker bees | 54M (queen 50M, workers 3M, links 1M) | queen 8, workers 4 | 8,192 tokens | 4096 | | |
| | `hive/BeeLLM-Hive-small.gguf` | The first hive. A small queen and her workers, with the older hive rules | 4.4M (queen 2.6M, workers 1.4M, links 0.5M) | queen 4, workers 4 | 1,024 tokens | 4096 | | |
| --- | |
| ## Tiny bees | |
| - **BeeLLM-30P**: the original bee. Her whole vocabulary is `bz`, `zz`, `zzz`, `bzzz`, `buzz`, a space, `~` and `!`. | |
| - **BeeLLM-42P**: more expressive. Adds `b`, `bu`, `z`, and `*stings you*`. | |
| - **BeeLLM-42P-Chaos**: chaos mode. Adds shouting: `BZ`, `BUZZ`, `BZZZ`… and `*STINGS YOU*`. | |
| The tiny bees (30P, 42P, 42P Chaos) are less chatty and reactive. They mostly just buzz. Give them a buzz or two and they keep buzzing. | |
| > **you:** `hello tiny bee` | |
| > **bee:** `buzz bzzzz buzzzzzzzzzzz buzz buzz!` | |
| > **you:** `buzz buzz` | |
| > **bee:** `bzzz bzzzzzzz bzzz buzz buzz buzz! bzzz bzzz~ bzzz bzzz buz buzz bzzzzzz bz! buzzzzz ` | |
| --- | |
| ## Mood bees | |
| - **BeeLLM-1M-Mood**: she hears what you say and answers with how she feels. | |
| - **BeeLLM-1M-Mood-Chaos**: the 1M Mood bee, but louder (`BRR`, `BZZT`), with more `*actions*`. | |
| - **BeeLLM-12M-Mood**: the biggest bee, with the widest range of buzzes and actions. She knows her name and holds grudges against objects and actions. | |
| - **BeeLLM-2.0-1M-Mood**: the 1M Mood bee, retrained on about 3,000 new conversations. More varied buzzes and actions, friendlier with gentle messages, and she has a favourite flower. Still a tiny bee, so she garbles a word now and then. | |
| - **BeeLLM-2.0-12M-Mood**: the 12M bee, retrained on the same new conversations. Everything the 12M could do, plus a richer voice, and a soft spot for orange flowers 🧡 | |
| - **BeeLLM-Scintilla-50M**: the hive queen on her own (8 layers, 10 heads). She has the mood tags, but her mood has momentum: after you're mean it climbs back one step at a time, and a real sorry is met sweetly. The biggest and most coherent Scintilla. | |
| > **you:** `*swats you*` | |
| > **bee:** `[angry:low] BZZzzT. swats? no. bad. bzzt. *stomps tiny feet*` | |
| > | |
| > **you:** `*gives you a flower*` | |
| > **bee:** `[neutral:low] hmm bz. flower? ...ok. thank you. *small nod*` | |
| > | |
| > **you:** `you're so pretty!!` | |
| > **bee:** `[loving:low] mmm~ exce? so pretty? me? bzz~ thank you~ you too~ *blushes*` | |
| > **you:** `*squishes you*` | |
| > **bee:** `[angry:high] BZZZZZT!! squishes me? NO! that was mean! ZZZZT *circles you*` | |
| > | |
| > **you:** `sorry, that was mean` | |
| > **bee:** `[loving:low] bzz~ sorry, you're good. I'm good. all good~` | |
| The Mood bees are expressive, have many more parameters, and fun to chat with. | |
| Every reply starts with a mood tag, based on how you treat her in the conversation: `[mood:level]` | |
| - **mood:** `neutral`, `happy`, `loving`, `sad`, `angry` or `confused` | |
| - **level:** `low` or `high` | |
| ### Mood reaction examples | |
| | say | she feels | | |
| |---|---| | |
| | `hi bee!` | neutral | | |
| | `*gives you a flower*` | happy | | |
| | `i love you` | loving | | |
| | `*sprays you with bug spray*` | sad | | |
| | `*swats you*` | angry | | |
| | `asdfghjkl` | confused | | |
| ### Making up with an angry bee | |
| Swatting her makes her angry. Being sweet to her straight after makes her angrier. She was trained to see through lovebombing. Say sorry. | |
| | after `*swats you*`, say | she feels | | |
| |---|---| | |
| | `sorry` | neutral | | |
| | `i love you` | angry (high) | | |
| | `*gives you a flower*` | angry (high) | | |
| ### Other things she does | |
| - **Put her somewhere** (`*puts you in my mouth*`, `*puts you in a jar*`, `*eats you*`) and she goes completely silent. Let her out (`*takes you out*`, `*opens the jar*`) and she comes back furious: | |
| > `*FILES A FORMAL COMPLAINT AGAINST THE JAR*` - (1M Mood Chaos) | |
| - **12M Mood knows her name.** It's Scintilla, and she lights up when you use it. `hi bee!` gets a polite `[neutral:low]`, but her name gets `happy:high` or `loving:high` every time: | |
| > **you:** `are you scintilla` | |
| > **bee:** `[loving:high] *perks up at her name* *cuddles your ear* mmm~ hmmbzz~` | |
| > | |
| > **you:** `what is your name` | |
| > **bee:** `[happy:high] *writes Scintilla in pollen* bzzzzz~! bzzbzz! brrr~` | |
| - **12M Mood holds grudges, against objects too.** Hurt her with something and she'll never forgive it: | |
| > **you:** `*gives you a flower*` | |
| > **bee:** `[happy:high] buzz buzz! bizz! bzzt! *hugs the flower like an old friend*` | |
| > | |
| > **you:** `*gives you a tomato*` | |
| > **bee:** `[happy:high] zip zip bzzbzz! *takes a tiny bite of the tomato*` | |
| > | |
| > **you:** `*throws the tomato at you*` | |
| > **bee:** `[angry:high] *STINGS YOU* *will never trust the tomato again* *attacks the tomato* BZZZZzzZ! BZZZZZZzZZZZ!!! BUZZZZzZzZz` | |
| - **Very rarely, she dies** (`*stings you and then dies*`). She stays silent for a few turns, then usually comes back with `*revives from the dead*`. Sometimes she doesn't. | |
| - **Ask her to sing** and she may not be in the mood: | |
| > `[sad:low] buzz. buzz...` - (12M Mood) | |
| --- | |
| ## Bee hives | |
| - **BeeLLM-Hive**: The neural bee hive. A 50M queen and up to 12 linked worker bees | |
| - **BeeLLM-Hive-small**: The first hive. A small queen and her workers, with the older hive rules | |
| - **Scintilla on her own**: the hive queen without her workers is `mood/BeeLLM-Scintilla-50M.gguf`. Unlike the hive files, she runs in llama.cpp | |
| The hive bees are not one bee. A hive is a queen and her workers, several small models running side by side that never read anothers words. | |
| They talk only in neuralese. Each bee reads the other bees' hidden states (the numbers inside their layers) through a learned link. | |
| That's how the whole hive knows what one bee felt. | |
| In short, it's a very tiny AI hive mind. | |
| The queen is **Scintilla**. The workers are smaller bees that share one brain, each with her own number and job: | |
| cleaner, nurse, builder, guard, or forager. | |
| ### What happens in a hive | |
| Every day, each bee is told only what happens to *her*, and writes her own line. Nothing she writes is shown to | |
| the others. They still act together: | |
| - **Swat a worker** and only she is told. Her sisters still come to guard her by number (`*guards bee 7*`), and | |
| Scintilla calls the guard. Swat again and a sister may sting you and die protecting her. | |
| - **A forager finds a flower** and does the waggle dance. The other foragers fly to that flower, though nobody named it to them. | |
| - **A sister hatches**, and the nurses feed her by her number. When a bee dies, an undertaker carries her out of the hive. | |
| - The workers' moods follow the queen's. | |
| **Cut the link** between them and all of that stops. Each bee still knows what happened to herself, but nothing | |
| else: | |
| | what the bees do | link on | link cut | | |
| |---|---|---| | |
| | sisters guard the bee you swatted | 100% | 0% | | |
| | foragers follow the dance to the right flower | 95% | 0% | | |
| | nurses feed the newborn by her number | 100% | 0% | | |
| | the queen calls the guard | 100% | 0% | | |
| | a swatted bee knows she was hurt (her own news) | 100% | 100% | | |
| (`BeeLLM-Hive.gguf`, hives of up to 6 workers, 256 test hives.) | |
| With the link cut, a hurt bee's sisters don't guard her. They write `*searches for queen* bzz? where`, as if the queen had gone. | |
| The full science write-up (how the link works, every number and its source, a look inside the bees' heads, and the limits) is in [`SCIENCE.md`](SCIENCE.md). | |
| # How to get started | |
| With [llama.cpp](https://github.com/ggml-org/llama.cpp): | |
| ``` | |
| llama-server -m mood/BeeLLM-1M-Mood.gguf -c 4096 --port 8081 | |
| ``` | |
| (or the name of whichever BeeLLM you downloaded) | |
| Then open `http://localhost:8081` and say hi. | |
| ## How to get started with the bee hives | |
| > **These don't run in llama.cpp.** A hive is several linked models in one file (architecture `beehive`), and | |
| > llama.cpp has no way to run models that read each other. You can visit them in your browser. | |
| > https://bee.pleroma.nexus/ (open **the hive**), with a small JavaScript runtime. How it all works is in | |
| > [`hive/README.md`](hive/README.md), if you want to set up your own architecture. | |
| However, The hive queen also works alone, as a 50M Scintilla for one-to-one chat: `mood/BeeLLM-Scintilla-50M.gguf`, and she runs in llama.cpp. | |
| She's the biggest Scintilla yet, with the most coherent replies. | |
| She has the mood tags of the Mood bees, and her mood climbs back one step at a time after you're mean to her: | |
| a flower right after a swat gets a careful `ok`, not a happy dance. A real sorry is always met sweetly. | |
| --- | |
| # Model details | |
| - **Developed by:** schizophyllume (skizy) | |
| - **Model type:** bee | |
| - **Architecture:** llama, F32 weights (the hive files are f16) | |
| - **Language(s):** buzz | |
| - **License:** Apache 2.0 | |
| ## Uses | |
| Buzzing and pollinating flowers | |
| ## Bias, risks, and limitations | |
| May sting | |
| ### Recommendations | |
| Give flowers to pollinate and cheerful words | |
| If you swat her, say sorry | |
| ## Training data | |
| Flowers, meadows, and nice forests |