01
Intake
Open or licensed model enters a versioned environment with explicit provenance.
LSB Tech / Model foundry From model weights
From model weights
to an artifact you control.
NeuroForge combines model intake, specialization with datasets, SFT and RL, compression and validation. The output is not an API dependency. It is a controlled artifact the NeuroGrid runtime executes inside infrastructure the organization owns.
MODEL → BINARY One controlled path.
One controlled path.
No runtime handoff.
02
Specialize
Domain data teaches the task, language and operating constraints that matter.
03
Compress
Distillation and quantization reduce footprint while preserving task quality.
04
Validate
Full-precision and compact artifacts face the same regression and quality gates.
05
Compile
Weights and metadata become a runtime artifact for CPU, Metal or CUDA.
06
Handoff
The artifact runs on the NeuroGrid runtime, a Go binary with no Python in the serving path.
Native by design
Training tools may use PyTorch. Production serving does not. This separation keeps experimentation flexible and deployment controlled.
Current state
Text inference is operational in the NeuroGrid runtime. On-demand ternary TTS has native CPU and CUDA paths. Realtime voice and additional modalities remain explicitly staged work.
Deployment envelope TARGET TARGET TARGET TARGET Examine measured evidence →
CPU
Offline and controlled workloads
Apple Metal
Local execution on Apple silicon
NVIDIA CUDA
Low-latency and throughput workloads
Edge / distributed
Field and multi-node topologies