KV-cache and memory-tier placement software, supplied with a service interface and a vLLM integration hook.
See what it solvesTHE PRODUCT, IN PRACTICE
SpinMem treats data placement across L2, HBM, CXL and NVMe as one supported optimization problem. It helps a team evaluate where inference data belongs rather than treating each tier’s settings independently. The supplied service interface and vLLM hook connect placement decisions to the serving workflow.
The cache competing for the fastest tier can exceed the capacity available to a serving workload.
SpinMem evaluates placement across the tiers represented in the deployment model, giving your system a deliberate allocation rather than an isolated infrastructure setting.
Rules chosen separately for different tiers can make the overall placement difficult to assess.
The application considers the supported placement problem together, so the team can review one placement decision against the modeled constraints and objective.
A useful plan is not enough when an engineer must manually translate it into the runtime.
The service interface and vLLM hook connect the placement decision to a supported serving workflow. Integration is focused on where the cache belongs, not on replacing the entire inference stack.
Placement decisions across the supported L2, HBM, CXL and NVMe model.
The supplied FastAPI service for accessing the placement capability.
The supplied vLLM hook for the qualified serving environment.
The modeled tiers, constraints and placement objective against which the application is evaluated.
SpinMem is fully available with CLI, MCP and headless access. Use the interface that fits your terminal, your agents or your automated workflow.
Run the product from a terminal or script, alongside the tools your team already uses.
Get CLI accessConnect MCP-compatible clients and agents to the product’s exposed capabilities.
Get MCP accessOperate without a graphical interface in unattended, scheduled or pipeline-driven work.
Set up headless accessProduct-specific inputs, hardware requirements and result guarantees still apply. An access mode does not change your license or deployment scope.
Explore all product interfacesA vLLM team wants to evaluate cache placement before committing to more infrastructure.
Illustrative customer use—not a new benchmark or completed customer deployment.Describe the supported tiers, capacities and cache-placement workload.
Connect the placement service to the qualified serving path.
Compare memory use and serving behavior with the existing placement policy.
Explore the product workflow, from your input to a result you can use.
KV-cache blocks compete for space across the available memory tiers.
Interactive explanation. This illustration does not execute customer workloads or display measured performance.
The product formulates supported placement choices as a QUBO. It is distinct from SpinCache representation/direct decode and KX MemoryForge offload/restoration. The best modeled placement is not a universal end-to-end speed or global-optimality guarantee.
$90,000/year is the retained standalone price proposal for its named scope. The platform plans buy qualified workflows and capacity, not a second charge every time a compatible method is selected. Existing bundle inclusions do not expand automatically.
Fully available with CLI, MCP and headless access. Runtime requirements, supported operations and license scope remain product-specific.
One named production deployment. Shared engine rights for this application included.
Scope & packagesBring a representative workload and the result your team needs. We’ll define the product, deployment and engineering scope around that task.
Activate this capability inside a qualified ai serving workflow. Compatible methods and required shared checking are included once within that workflow; the product remains independently purchasable.
Capacity, release, environment, result contract and operating responsibilities are specified before activation. Choosing an additional method is not permission to add an unrelated workload or unqualified deployment.
Explore AI Serving Choose workflow capacityCommission the model, integration and acceptance work needed for your task. Existing products stay independently available.
Use the software directly or ask us to scope the implementation.
One capability. A connected workflow.
A team to put it into practice.