HivenetQuant

HivenetQuant publishes optimized versions of open models for efficient inference.

Our current work focuses on quantization: reducing the memory and compute required to run models while preserving model quality as closely as possible. Each release documents the optimization method, target hardware, benchmark results, and tested serving configuration so developers can evaluate the trade-offs for their own workloads.

What we publish

HivenetQuant repositories include:

• optimized and quantized model checkpoints
• comparisons with an appropriate higher-precision baseline
• benchmark results and testing methodology
• tested inference and serving configurations
• details about the quantization method and target hardware
• instructions for running and evaluating the model

Requirements vary between releases, so check the individual model card before using a checkpoint.

Our current releases include NVFP4 quantizations designed for NVIDIA Blackwell GPUs.

How we evaluate our models

Optimization involves trade-offs between memory use, inference performance, hardware support, and model quality.

We publish the configuration used for our benchmarks alongside the results wherever possible. Performance figures should be considered together with the hardware, runtime, workload, sequence length, batch size, and other conditions under which they were measured.

When different parts of a model use different precision levels, we document that distinction rather than describing the checkpoint as a uniform quantization.

Using our models

Each repository contains the requirements and serving instructions for that release, including supported runtimes where applicable.

Start with the model card for the checkpoint you want to use.

About HivenetQuant

HivenetQuant is Hivenet's Hugging Face organization for optimized open models.

Hivenet provides GPU compute and infrastructure for running and serving AI workloads.