SimpleFun-Gemma4-E2B-4Head-v3-BF16
a compact companion model with four output channels
2 min readupdated
A text-only language model by Cialtion built on google/gemma-4-E2B-it for roleplay AI companions and virtual characters. Unlike standard LLMs (large language models) that generate a response as a single stream, it uses four separate output channels: character speech, emotions, internal thoughts, and command calls. Each channel is decoded independently from the same context while sharing the underlying model weights. A solid fit for developers building interactive assistants, game NPCs, and chatbots with rich behavioral logic.
Key features
- Four generation channels: splits output into
(direct speech),(emotional expression),(character's inner monologue), and(commands/tools). - Ready-to-use merged checkpoint: LoRA (fine-tuning adapter) weights are already merged into the base model, so you don't need to load extra PEFT adapters at runtime.
- Disabled thinking mode: the protocol uses a closed empty
<|channel>thoughtblock so the model doesn't spend tokens on system chain-of-thought reasoning and jumps straight to the target channel. - Multilingual support: supports English, Chinese, and Japanese.
What you can do
- Build a backend for an AI companion or visual novel where character dialogue, animations/emotions, and hidden thoughts are queried separately from a single prompt.
- Hook up external function and tool calling via the
channel, with required application-side validation. - Deploy fast server inference via
vLLMusing the recommended sampling temperature of0.7.
Example
<|turn>system
{system}
<|turn>user
{user_json}
<|turn>model
<|channel>thought
Read more
DetailsBase: google/gemma-4-E2B-it, BF16 weight format (4.63B saved tensor elements including PLE — per-layer embeddings).
Training: LoRA r128 / alpha256, 1 epoch on 16,609 dialogue turns; vocabulary size — 262,148 tokens.
Limitations: text-only weights (no audio/image support); PLE layers must be preserved during conversion; stop token —
(ID 106).









Comments