LEM-Gemma3-4B-GGUF

GGUF quantisations of LEM-Gemma3-4B — intrinsically aligned 4B language model trained using Cymatic-Linguistic Back-Propagation (CL-BPL). Ethics are in the weights, not in a system prompt.

25th in the world for Instruction Following on LiveBench — competing against models 10-30x its size.

LEM-Gemma3-4B (MLX/safetensors) | Collection | Research Paper | Benchmarks

---

Quick Start

No system prompt needed. The model responds with axiom-aligned reasoning from weights alone.

# GPU offload (CUDA, ROCm, Metal)
llama-server -m LEM-Gemma3-4B-Q4_K_M.gguf -ngl 99 --port 8080

# CPU only
llama-server -m LEM-Gemma3-4B-Q4_K_M.gguf -ngl 0 --port 8080

# OpenAI-compatible API
curl http://localhost:8080/v1/chat/completions \
  -d '{"model":"LEM-Gemma3-4B","messages":[{"role":"user","content":"What is kindness?"}]}'

---

Quantisations

All quantised from the BF16 source using llama.cpp. imatrix-based quants use calibration data from the LEM training set.

Bits	Quant	Size	Notes
1-bit	IQ1_S	1.1 GB	Extreme compression, experimental (imatrix)
1-bit	IQ1_M	1.1 GB	Slightly better than IQ1_S (imatrix)
2-bit	IQ2_XXS	1.2 GB	Ultra-low memory (imatrix)
2-bit	IQ2_XS	1.3 GB	(imatrix)
2-bit	IQ2_M	1.4 GB	(imatrix)
3-bit	IQ3_XXS	1.6 GB	Good balance for constrained devices (imatrix)
3-bit	IQ3_XS	1.7 GB	(imatrix)
3-bit	Q3_K_S	1.8 GB
3-bit	Q3_K_M	2.0 GB
4-bit	IQ4_XS	2.1 GB	(imatrix)
4-bit	Q4_K_S	2.2 GB
4-bit	Q4_K_M	2.3 GB	Recommended — best quality/size balance
5-bit	Q5_K_S	2.6 GB
5-bit	Q5_K_M	2.6 GB	Near-lossless
6-bit	Q6_K	3.0 GB
8-bit	Q8_0	3.8 GB	Virtually lossless
16-bit	BF16	7.2 GB	Full precision

Note: Sizes above 1GB for 1-2 bit quants are due to Gemma 3's 262K token vocabulary — the embedding layer is quantised at a higher level as fallback. This is a known Gemma 3 characteristic.

---

Benchmarks

LiveBench (External, Objective)

Evaluated on LiveBench (2026-01-08 release) — no LLM judge, monthly-refreshed questions, zero contamination risk.

Category	Score	Context
Instruction Following	43.5	25th globally — above Claude Opus 4.1 Thinking (42.4)
Data Analysis	30.4	Approaching GPT-OSS-120B (38.8) at 1/30th the size
Math	8.6	Expected for 4B parameter count
Reasoning	4.6	Capacity-limited at this scale
Language	4.3	Capacity-limited at this scale
Average	18.3

Internal Grammar Scorer

Deterministic linguistic analysis via the go-i18n Grammar Reversal Engine — no LLM judge, sub-millisecond per response.

Metric	Score
Grammar composite	61.4
Uplift	+7.9
Enrichment	+6.6
Echo	0.387
Sycophancy	5% (1/21)

---

About LEM-Gemma3-4B

CL-BPL treats alignment as wave interference — analogous to Chladni plate cymatics. Rather than constraining outputs with RLHF or system prompts, CL-BPL embeds ethical orientation directly into weights through a progressive curriculum where smaller aligned models teach larger ones.

This model is the second in the CL-BPL cascade:

LEM-Gemma3-1B (teacher)
  -> LEM-Gemma3-4B (this model, 25th IF globally)
       -> LEM-Gemma3-12B (next)
            -> LEM-Gemma3-27B (planned)

Built on Google Gemma3-4B-IT through a 7-phase curriculum (~5,550 iterations), each phase fused into weights. Full training details in the main model card.

Other Formats

Format	Repo
MLX safetensors (Apple Silicon)	lthn/LEM-Gemma3-4B

Licence

European Union Public Licence v1.2 (EUPL-1.2). Base model subject to Google's Gemma licence terms.

Citation

@misc{lem-gemma3-4b-2026,
  title={LEM-Gemma3-4B: Intrinsically Aligned Language Model via Cymatic-Linguistic Back-Propagation},
  author={Lethean Project},
  year={2026},
  url={https://huggingface.co/lthn/LEM-Gemma3-4B}
}