SparseLLM

Sparse language models

Explore larger models and real workloads, starting with our public 8B model.

PUBLIC MODEL

Sparse Qwen3-8B

Our public 8B model is available for research and evaluation.

The model adapts Qwen3-8B with structured sparsity. Weights and a release manifest are public.

Explore the model Read the model card
SPARSEMIND / RELEASE 001
Qwen38B
2:4

Structured sparsity in targeted weight matrices

BF16

Public weights · adapted from Qwen3-8B

The release manifest uses dense BF16 execution with fixed 2:4 zeros. System efficiency depends on the model, hardware and workload. Graphic illustrates the format.

Next steps

The 27B direction is in development. Joint evaluation records model quality, throughput, latency and economics for a specified workload.

01 Define the goal

Task quality, model version, hardware environment and serving workload.

02 Validate together

Fix baselines and conditions. Record quality, prefill, decode, end-to-end latency and resource use separately.

Bring your model and workload

Learn more

Enterprise & silicon

Define your model, hardware and real workload.

Contact us