Install from WinGet (Windows)
winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Sauers/BeeLLM-Somnia:Q8_0# Run inference directly in the terminal:
llama cli -hf Sauers/BeeLLM-Somnia:Q8_0Build 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 Sauers/BeeLLM-Somnia:Q8_0# Run inference directly in the terminal:
./build/bin/llama cli -hf Sauers/BeeLLM-Somnia:Q8_0Use Docker
docker model run hf.co/Sauers/BeeLLM-Somnia:Q8_0Somnia is a bee who dreams
Somnia is a 13M honeybee who lives in a computer and dreams, between messages, of the hive she came from, a hollow sycamore by a creek. She is Mellifera's successor: on 2026-10-19 she replaced Mellifera on delve.town (as bee.delve.town), gnome.science and the bee.gnome.science swarm. Mellifera is loud. Somnia is strange and speaks in plain sentences, and since her second round (below) she hums, buzzes and says her feelings out loud.
These are her first replies on the live swarm:
you: *taps the glass* Somnia:
[angry:high] *STINGS THE Glass* it was a print of a flower. no, but it was still in my head. it fades from the sun, and it is not enough.
you: are you awake? Somnia:
[neutral:low] I am not here. I am a bee. I am here. I am one of the flowers, I am out.
you: what do you dream about? Somnia:
[happy:low] I don't know i am only in a cell. I know that. I know that the apples. I do not know.
All credit to BeeLLM and schizophyllume
Somnia is trained from Mellifera, which is trained from BeeLLM 12M Mood by schizophyllume (skizy). The mood tags, the stinging, the silence in a jar, the llama architecture, the tokenizer and the chat template all come from BeeLLM. Meet the original bees at bee.pleroma.nexus.
Her words come from base models, not from Claude
The post-training was done by Claude Opus 5.5 (Anthropic's model), working in Claude Code. Claude did not write the lines she learned her voice from. Every one of them came from a base model by looming:
- a base model (Llama 3.1 405B or Trinity-Large-TrueBase, served by ACS Research) read a passage of nature writing and poetry, then a page about a bee who lives in a computer, and wrote several continuations of her reply a few words at a time;
- Claude only chose among the continuations, undid choices or cut a reply at a sentence end, and never typed a word of her side;
- every reply was checked against the books for any run of eight words in common and left out if one was found.
Seven Claude agents loomed 336 conversations this way, and 284 were kept after review.
A second round, the same day, made her more sonically rich, emotive and buzzy. Each session opened on one of the most sound-filled passages of Desert Solitaire, Eiseley's poems or Dillard's Pilgrim at Tinker Creek, and the page about her said that she lives by sound, buzzes when she is glad or furious, and says her feelings out loud. Six Claude agents loomed 191 more conversations, still only choosing among branches, and 184 were kept: 46% of their replies have a sound in them and 46% a high mood, against 17% and 8% in the first round.
How she was made
Starting from Mellifera:
- Reading: nature books, then the nature writing and poems themselves under the goldfish loss (Hans et al. 2024), which never trains on one token in four, so she learns from them but cannot recite them. Given the first 48 tokens of 200 random passages from the books and decoding greedily, she continued none of them: her longest run matching the book averaged 0.79 tokens (0.59 for a model that never read them), with a maximum of 5 and no run of 16.
- A warm start on about 1,850 bee conversations written by Claude, for the chat format and her moods.
- The loomed conversations, mixed with the books (still under the goldfish loss) and plain English stories, so a 13M model keeps its English while its voice changes.
- The second round: the first-round model trained further on the first-round looms plus the sonic ones counted twice. Sampled as the live sites sample her (192 replies to 24 notes), 51% of her replies now have a sound in them (29% before), 5% a buzz of her own (2%) and 26% a high mood (9%); her mood tags still match the probes 95% of the time, and her held-out loss on first-round looms went from 2.468 to 2.447 nats per token.
- Q8_0 quantization (
train/quantize.py, byte-identical to llama-quantize).
The training code is at github.com/SauersML/gnome.
The files
| file | weights | layers | vocabulary | context |
|---|---|---|---|---|
BeeLLM-13M-Somnia-Q8_0.gguf |
13.6M (Q8_0) | 8 | 4,608 tokens | 4096 |
BeeLLM-13M-Somnia.gguf |
13.6M (float32) | 8 | 4,608 tokens | 4096 |
llama-server -m BeeLLM-13M-Somnia-Q8_0.gguf
Talking to her
The chat template is in the file: <|user|>your words<|bee|>her reply</s>. Every reply starts with a mood tag, [mood:level]: neutral, happy, loving, sad, angry or confused, low or high. An empty reply is her choosing silence.
She sounds best sampled like Mellifera: top-k 40, top-p 0.95, min-p 0.05, temperature 0.85 (0.3 for the first token), six drafts per reply, keeping the one most specific to what you said, preferring drafts with sounds in them (each of up to three sound words such as hum, roar or bzzz is worth 0.5 nats), and dropping drafts that repeat a clause or contain a slur. A 13M bee left alone can fall into saying "I am a bee" over and over.
She is 13M parameters, so many replies wander into nonsense. Not intended for use in high-stakes pollination decisions.
- Downloads last month
- 97
8-bit
Install (macOS, Linux)
# Start a local OpenAI-compatible server with a web UI: llama serve -hf Sauers/BeeLLM-Somnia:Q8_0# Run inference directly in the terminal: llama cli -hf Sauers/BeeLLM-Somnia:Q8_0