GoingMerry is the next product we are building after ICE.
The Core Trinity
Native kernels
Custom CUDA, Metal and AVX kernels for local inference.
Local by default
Runs on your own hardware, including air-gapped and on-prem deployments.
Several hardware targets
Kernels for NVIDIA GPUs, Apple Silicon and x86 CPUs.
Benchmark Metrics
Internal measurement. Methodology not yet published.
Under the Hood.
GoingMerry pairs a Go orchestration layer with a C++ tensor inference engine, with custom kernels for each hardware target.
- Layer 01
Go Orchestrator
Async Scheduler & Memory Broker - Layer 02
C/C++ Tensor Engine
Hardware Compiler & L3 Cache Control - Layer 03
CUDA / Apple Metal / AVX
Local operations
Hardware-Aware Toolchain
Custom C/C++ compiler toolchain implementing branchless hot-path optimization, AVX-512/NEON vectorization on CPUs, and tensor core instructions on GPUs.
Mastering Memory Latency
Cache-aware runtime that organizes model weights to align with CPU pre-fetching mechanisms, establishing high-bandwidth L3 cache tunnels.
Predictive Performance
An empirical forecasting model that predicts local tokens-per-second output based on target memory bandwidth and model parameter sizes.
Go + C++ Hybrid Engine
A polished, high-level Go orchestration layer for API and network logic driving a low-level C++ tensor inference engine compiled for speed.
Simple Licensing
Start free. Scale when you need to.
Community
Free for individuals. Run models locally.
- Free local execution
- Native CPU and GPU kernels
- Air-gapped privacy
Enterprise
For teams needing compliance, support, and scale.
- SSO, RBAC & audit logs
- Fleet management & Kubernetes
- Air-gapped deployment
Up and running soon.
We are calibrating local compiler targets and model sharding frameworks. Join the waitlist to get notified the second GoingMerry becomes available for local installation.
