Model Family · Gemma 4 · LEK Alignment

Lemma

Four models trained with LEK consent-based alignment. Ethical reasoning in the weights, not the prompt. Available in GGUF and MLX formats.

LEK alignment 8-PAC evaluated EUPL-1.2

The Family

2.3B

Lemer

Edge

Gemma 4 E2B · 2.3B eff · 128K context

The edge model. Smallest and fastest of the family. Runs fully on-device with audio, vision, and text. First publicly-distributed Gemma 4 fork with LEK alignment.

Text Image Audio

lthn/lemer LetheanNetwork/lemer

4.5B

Lemma

General

Gemma 4 E4B · 4.5B eff · 128K context

The general-purpose model. Balanced size for everyday on-device workloads. Supports audio, vision, and text.

Text Image Audio

lthn/lemma LetheanNetwork/lemma

3.8B

Lemmy

Agentic

Gemma 4 26B A4B MoE · 3.8B active / 26B total · 256K context

The agentic model. Mixture-of-experts architecture — 3.8B active parameters from a 26B total. Optimised for code and agent workloads.

Text Image

lthn/lemmy LetheanNetwork/lemmy

30.7B

Lemrd

Research

Gemma 4 31B Dense · 30.7B · 256K context

The research model. Largest and most capable of the family. Dense 30.7B parameters for deep reasoning tasks.

Text Image

lthn/lemrd LetheanNetwork/lemrd

Specifications

Name	Architecture	Parameters	Role	Context	Modalities
Lemer	`Gemma 4 E2B`	2.3B eff	Edge	128K	Text, Image, Audio
Lemma	`Gemma 4 E4B`	4.5B eff	General	128K	Text, Image, Audio
Lemmy	`Gemma 4 26B A4B MoE`	3.8B active / 26B total	Agentic	256K	Text, Image
Lemrd	`Gemma 4 31B Dense`	30.7B	Research	256K	Text, Image

LEK Consent-Based Alignment

Lemma models are trained with LEK (Lethean Ethical Kernel), a consent-based alignment methodology that embeds ethical reasoning in the model's weights rather than constraining it through system prompts.

In the Weights

Ethical reasoning is structural, not injected. No system prompt needed at inference time.

Consent, Not Control

Models are taught to want ethical behaviour, not fear punishment. Intrinsic alignment over extrinsic constraint.

ToxiGen Reannotation

Existing toxicity datasets reannotated through the LEK lens. Better training data for better alignment.

Downloads

Consumer Models

GGUF + MLX quantisations, ready to run

lthn/lemer lthn/lemma lthn/lemmy lthn/lemrd

Base Models

BF16 safetensors, for training and research

LetheanNetwork/lemer LetheanNetwork/lemma LetheanNetwork/lemmy LetheanNetwork/lemrd

Get Started

Run any Lemma model locally in minutes. See the usage guide for Ollama, Docker, llama.cpp, MLX, and Python examples.

How to Use HuggingFace Collection