MODEL STORE
De-Risked releases. One-time licenses.
Buy an available Blackfrost model-weight release outright for a one-time price. Each purchase grants licensed access to the exact gated Hugging Face repository shown on its model card. Runtime tooling, compute, hosting, installation, and custom engineering are separate. Upstream licenses, export controls, sanctions, and local law still apply.
Exact weights.
Clear boundaries.
[ MODEL RELEASES ]
Every release has a one-time price, named repository, exact package footprint, verified revision, selection guidance, and deployment boundary. Review the release card and license before paying securely on Stripe.
The full 753B BF16 reference master for evaluation and downstream conversion.
The complete 256-expert GLM-5.3 NVFP4 package with a documented 8× NVIDIA B200 target.
The 2.4T-parameter reference master for evaluation and custom conversion.
A two-node deployment tier between the BF16 master and W4A4 release.
The two-node W4A4 NVFP4 deployment release of the 2.4T MoE.
The complete native Kimi K3 release with all 896 routed experts retained.
Bit-exact MXFP4 experts packaged for a single 8×B200 node.
The full-precision GLM-5.2 reference master for evaluation and conversion.
Block-FP8 GLM-5.2 packaged for a single eight-GPU node.
The full 256-expert GLM-5.2 research build in NVFP4.
A smaller research tier pruned under non-uniform expert budgets.
Prices are in USD. A purchase licenses the named model-weight repository; it does not transfer upstream intellectual property or include inference tooling, an API, compute, hosting, installation, custom engineering, ongoing operations, or future catalog releases. Shadow CLI is a separate open-source Community project, not part of a model purchase.
[ CUSTOM HOSTING · SEPARATE ENGAGEMENT ]
Blackfrost can separately scope capacity and topology, runtime and quant selection, containerized endpoints, autoscaling and routing, observability, security controls, rollback, load testing, and customer-cloud handoff. The architecture, implementation, and any ongoing operations or SLA are quoted for the actual workload; none are bundled with a catalog model.
Plan scalable hosting