It is trained with PoSTMEDIA's in-house Capability-Preserving Full Fine-Tuning recipe — a full-parameter SFT pipeline designed so that customization does not erode the reasoning, instruction-following, and multilingual abilities of the Gemma-4 base model.

Highlights

Full fine-tuning of Gemma-4-26B-A4B (MoE, ~4B active) — not LoRA, not an adapter
Base capability preserved — pretraining knowledge and reasoning skills remain intact after SFT
Dataset-flexible — any combination of curated instruction / domain / persona datasets can be composed into a single full-FT run
Efficient inference — MoE architecture keeps active parameters low at serving time

Model Overview

Specification	Details
Base Model	`google/gemma-4-26B-A4B-it`
Parameters	26B total / ~4B active
Architecture	Decoder-only Transformer (MoE)
Training Precision	BF16
Inference Precision	BF16
Context Length	Inherits from Gemma-4 base
Fine-Tuning Method	Full-parameter SFT (Capability-Preserving recipe)

Capability-Preserving Full Fine-Tuning

Naive full fine-tuning of large pretrained LLMs often damages the base model's general abilities — a well-known trade-off when SFT is pushed too far. PoSTMEDIA's recipe is built specifically to avoid this.

For Lux-V1, three design choices keep the Gemma-4 base intact while still allowing deep adaptation:

Selective trainable modules for MoE. Only the attention projections (q_proj, k_proj, v_proj, o_proj) are updated. Expert FFN layers are left frozen, which prevents the sparse routing structure of Gemma-4-26B-A4B from collapsing during SFT.
Architecture-tuned learning rate. The LR is calibrated for MoE training dynamics to keep optimization in the stable regime where the base distribution is preserved.
Continuous base-capability evaluation. Evaluation runs at the start of training and at every epoch, so any regression in base-model quality is caught early rather than discovered post-hoc.

This means Lux-V1 can be re-trained from the same base with arbitrary mixtures of datasets — identity, domain knowledge, instruction-style, reasoning — without losing what Gemma-4 already knows.

Training Configuration

Parameter	Value
Fine-Tuning Method	Full-parameter SFT (attention-only, MoE-safe)
Precision	BF16
Distributed Strategy	DeepSpeed ZeRO-3 + CPU offload
Training Infrastructure	NVIDIA H200 × 8

Quick Start

pip install transformers accelerate

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "PoSTMEDIA/Lux-V1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

prompt = "Explain why preserving base-model capability matters during fine-tuning."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Use Cases

Enterprise assistants requiring domain adaptation without losing general reasoning
Persona / identity-aligned chat applications
Instruction-tuned downstream tasks built on a Gemma-4 foundation
Efficient on-premise serving where MoE active-parameter cost matters

Safety & Limitations

Inherits the safety characteristics of the Gemma-4 base; output guardrails are recommended for production.
Not intended for medical, legal, or financial decision-making.
May occasionally hallucinate — human review is recommended for critical outputs.

Citation

bibtex

@misc{lux_v1_2026,
  title  = {Lux-V1: Capability-Preserving Full Fine-Tuning of Gemma-4-26B-A4B},
  author = {PoSTMEDIA AI Lab},
  year   = {2026},
  publisher = {Hugging Face}
}

Contact

PoSTMEDIA AI Lab