OPEN MODEL · OPEN WEIGHTS

Open sparse models

SparseMind’s first open-weight sparse large language model. Built on Qwen3-8B, it turns an original scientific approach into a model that developers can download and evaluate.

8B
Qwen3-8B2:4BF16

Download. Evaluate. Explore.

Selected weight matrices use a 2:4 sparse pattern. The model card provides release details, execution instructions and conditions for developers to build their own evaluations and experiments.

The release documentation specifies dense BF16 matrix multiplication (GEMM). Any performance gains from sparse kernels must be validated separately on the target hardware and workload.

Open on Hugging Face

A model milestone.
A longer scientific journey.

This release makes structural learning accessible through a model that others can explore directly. Quality, sparse structure and target hardware remain central to its path toward applications.

Discover Cannistraci–Hebb Training