PoSTMEDIA/Lux-V1-Pro

VISIONConcurrency Cost:2Model Size:31BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:May 18, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

PoSTMEDIA/Lux-V1-Pro is a 31 billion parameter fully fine-tuned large language model developed by PoSTMEDIA AI Lab, built upon google/gemma-4-31B-it. It utilizes a Capability-Preserving Full Fine-Tuning recipe to ensure that customization does not degrade the base model's reasoning, instruction-following, and multilingual abilities. This dense model is designed for maximum capability in demanding downstream tasks, preserving pretraining knowledge and general reasoning skills. It is particularly suited for high-capability enterprise assistants, domain-specialized models, and persona-aligned chat applications requiring strong general-purpose abilities.

Loading preview...

Lux-V1-Pro: Capability-Preserving Fine-Tuning

Lux-V1-Pro is a 31 billion parameter fully fine-tuned large language model from PoSTMEDIA AI Lab, based on google/gemma-4-31B-it. It stands out due to its Capability-Preserving Full Fine-Tuning recipe, a method designed to deeply customize the model without eroding the foundational reasoning, instruction-following, and multilingual skills of the Gemma-4 base.

Key Capabilities & Features

  • Full-parameter fine-tuning: Every weight of the 31B dense Gemma-4 base is updated, targeting maximum capability.
  • Base capability preserved: PoSTMEDIA's unique fine-tuning approach prevents catastrophic forgetting, maintaining the base model's pretraining knowledge and reasoning skills.
  • Dataset-flexible: Supports arbitrary combinations of instruction, domain, and persona datasets for fine-tuning without compromising general abilities.
  • Optimized for demanding workloads: Intended for complex reasoning and generation tasks where a larger, dense backbone is beneficial.

Differentiators

Unlike naive full fine-tuning that can degrade general abilities, Lux-V1-Pro employs specific design choices:

  • Conservative parameter updates: All parameters are trainable but under a tightly controlled optimization regime.
  • Architecture-tuned learning rate: A calibrated lower learning rate for the 31B backbone avoids the catastrophic-forgetting common in aggressive full fine-tuning.
  • Continuous evaluation: Base-model quality is evaluated throughout training to catch any regression early.

Use Cases

This model is ideal for:

  • Developing high-capability enterprise assistants and reasoning agents.
  • Creating domain-specialized models that must retain strong general-purpose abilities.
  • Implementing persona/identity-aligned chat with robust instruction following.