LUMINIUM Ultimate Cube — 425M
A 425.8M parameter hybrid convolutional-attention language model built through layer surgery, collective-consciousness distillation, and cognitive-cube steering — distilled from an 8-model GPU cluster.
At a glance
Parameters
425.8M
expanded from 354.5M
Architecture
12 conv + 8 GQA
20 layers total
Context
128K
tokens
Speed (Q5_K_M)
164 tok/s
AMD Radeon VII
Speed (bf16)
128 tok/s
AMD Radeon VII
Domain pass
16 / 16
math · logic · code · safety · …
Overview
Model ID
mambiux/Luminium-Gixel-Cube-v1
Author
mambiux (with Claude Opus 4.6)
Base model
LiquidAI/LFM2-350M
License
Apache 2.0
Pipeline
text-generation
Languages
10
Key features
- Hybrid conv-attention architecture
- 128K context window
- Layer surgery (DARE+TIES merging)
- Cognitive-cube steering
- Multi-turn coherence
- 28–64% faster than base
- Natural conversational tone
- Self-awareness signals
Architecture & surgery
Expanded from LFM2-350M (16 → 20 layers) using cross-model architectural surgery. Four new layers were inserted via DARE + TIES merging, then the result was fine-tuned on a 45-source balanced curriculum distilled from an 8-model GPU cluster. Cognitive-cube steering positions the model in a 3D cognitive space, with inverse-distance weighting toward 8 specialist corner models.
Cognitive Cube — 8 corner models
| Corner | Model | Params | Role |
|---|---|---|---|
| fwd · dexo · up | Server 1 (Qwen3-80B) | 80B | predict — structured · abstract |
| fwd · dexo · down | Server 2 (Qwen3.6 MoE) | 35B | predict — structured · concrete |
| fwd · levo · up | Server 3 (granite-4.1) | 8B | predict — creative · abstract |
| fwd · levo · down | Server 4 (LFM2-12B) | 12B | predict — creative · concrete |
| back · dexo · up | Server 5 (Qwen3.6-A3B) | 35B | reflect — structured · abstract |
| back · dexo · down | Server 6 (Qwen3.5-9B) | 9B | reflect — structured · concrete |
| back · levo · up | Server 7 (lumina-lexiR1-8B) | 8B | reflect — creative · abstract |
| back · levo · down | Server 8 (Berthier-24B) | 24B | reflect — creative · concrete |
Available formats
| Format | Size | Best for |
|---|---|---|
| model.safetensors (bf16) | 1.6 GB | Fine-tuning · research |
| LUMINIUM-ULTIMATE-CUBE.gguf (bf16) | 815 MB | Full-precision inference |
| LUMINIUM-ULTIMATE-CUBE-Q5_K_M.gguf | 297 MB | Production · edge devices |
Training
Method
LoRA (PEFT) fine-tuning
LoRA rank / α
16 / 32
Learning rate
3e-5
Epochs
3
Batch size
8 (effective 16, grad-accum)
Max seq length
768
Precision
bfloat16
Hardware
AMD Radeon VII (16 GB HBM2, ROCm 6.2)
Training data — 18,593 records · 45 sources
- General instruction — 3,499
- Reasoning & CoT — 5,995
- Agentic & tool use — 2,573
- Code — 599
- Knowledge — 2,495
- Feedback & editing — 1,341
- Analytical reasoning — 900
- + 11 specialized categories
Usage
Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer import torch model = AutoModelForCausalLM.from_pretrained( "mambiux/Luminium-Gixel-Cube-v1", torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True, ) tokenizer = AutoTokenizer.from_pretrained( "mambiux/Luminium-Gixel-Cube-v1", trust_remote_code=True, )
llama.cpp
llama-server -m LUMINIUM-ULTIMATE-CUBE-Q5_K_M.gguf \ --host 0.0.0.0 --port 8877 -c 4096 -ngl 99
Limitations
- Inconsistent self-identification
- Basic routing classification
- Occasionally verbose on simple Q&A
- Edge-case logic drift in quantized version
- Requires trust_remote_code=True
- Conv layers not supported by all backends
Citation
BibTeX
@misc{luminium2026,
title = {LUMINIUM ULTIMATE CUBE: Cognitive Cube Steering and Collective
Consciousness Distillation for Small Language Models},
author = {mbx and Claude Opus 4.6},
year = {2026},
note = {Built on LiquidAI/LFM2-350M with layer surgery, 8-model
collective distillation, and geometric cognitive steering}
}