Pantry is free and open source. The rest of the AI stack (Iron, Butter, Syrup, Whisk, Waffle) is in private early access with pricing to be determined.
Key features
Constant-state attention: models serve any context length from the same memory per request
HCGM: models run from disk storage instead of memory
M-Rail: context processed by a model can be reused instantly in future requests
CPUDA/sGPU: run unmodified CUDA programs on CPUs with bit-identical results
Organ grafting: transfer capabilities between models without training
Personal instruction set (ISA) compiled on device from phone signals
Population ISA: compile laws from many people without personal records
Digital twins of customers, fleets, cities or markets
Results identical across all machines and architectures
Computation cost follows novelty not data size
Pantry: free, open-source vulnerability scanner for macOS and Linux
Iron: GPU kernels optimized for Apple silicon, NVIDIA CUDA, and AMD ROCm
Butter: multi-modal inference engine for text, vision and audio
Whisk: memory and context management to mix documents and context into requests
Syrup: control plane for monitoring, telemetry, access control and cost per GPU
Waffle: agent harness for building assistants
Runs on Apple silicon, NVIDIA RTX, H100, H200, B100, B200 and AMD ROCm
Local execution keeps data on your hardware by default
What makes it different
Constant memory cost regardless of context length or data size
Exact reuse: every computed result is kept and reused instantly
Portable execution with bit-identical results across all hardware
Models run from storage without retraining or fitting in memory
Privacy-preserving: raw data never leaves the device
Built for lower cost per token through efficiency optimization
Private by default; data stays on your hardware
Supports multiple hardware platforms with write-once kernels
Pantry offers free, open-source vulnerability scanning and patching
Enables AI on commodity hardware people already own