---
title: Beatrix — AlephLLM Chat
emoji: 🐠
colorFrom: yellow
colorTo: yellow
sdk: gradio
sdk_version: 6.23.1
python_version: '3.12'
app_file: app.py
pinned: false
license: mit
short_description: Two byte-level AlephLLM crafts — arms and full splat
---

# Beatrix — AlephLLM chat

Two lineages of the same byte-level architecture, served side by side.
Pick the **Craft** first; everything below it follows.

**mini-beatrix-1** — 112.5M parameters, pretrained on 15.3 billion raw
bytes, serving with a **library of detachable arms**: small aleph
adapters (3.2M–5.3M parameters), each trained on one exact frozen core.

**mini-beatrix-2s** — 237.1M parameters, **full splat**: the signed
address read runs in *every* block (v1 used it in three), governed from
birth so the anchors never crowd. The active mission — still
pretraining, no arms yet, and not yet through the phase that teaches a
conversation frame, so **Completion** is the honest view of it.

- **Chat** — conversation, in whichever frame the selected craft and arm
  were actually trained on.
- **Completion** — raw next-byte continuation.
- **Model switch** — flip to *Core only* and the arm detaches
  **bit-for-bit**: you are talking to the untouched pretrained base.
  That exactness is the architecture's central claim, and this is where
  you can check it.
- **KV cache toggle** — cached incremental decode vs full recompute,
  with live bytes/s. The aleph attention layers keep a *constant-size*
  decode state regardless of context length.

**The specials format (2s).** Thirteen structural tokens live on the
byte values UTF-8 can never produce — `⟦DOC⟧` ends every document,
`⟦SYS⟧ ⟦USER⟧ ⟦MODEL⟧ ⟦END⟧` frame a conversation. Because encoded text
cannot contain them, nothing typed into the box can forge the frame: the
guarantee is arithmetic, not convention. They render visibly here rather
than as mojibake.

Weights, arms, manifests and TensorBoard logs live in
[alephllm-mini-beatrix-training](https://huggingface.co/AbstractPhil/alephllm-mini-beatrix-training);
conversations are written (with notice) to this Space's persistent
storage and periodically published to
[alephllm-chat-history](https://huggingface.co/datasets/AbstractPhil/alephllm-chat-history).

They are small and early: conversational in shape, thin on knowledge, and
confidently wrong at times. Short exchanges suit them best.

Code: [alephllm](https://github.com/AbstractEyes/alephllm) ·
adapters: [amoe-lora](https://github.com/AbstractEyes/amoe-lora)