Roadmap
Where we're headed
Phase 1: Precision & Performance
Float16 & BFloat16 Support
In ProgressExtend Candy with half-precision and brain floating-point support for memory-efficient training and inference.
- float16 (IEEE 754 half-precision)
- bfloat16 (Brain floating-point)
- Mixed-precision training utilities
- Automatic type casting
Phase 2: Inference & Interoperability
ONNX Bindings
PlannedONNX ecosystem bindings for Go — load, run, and optimize ONNX models natively.
- ONNX model parser and loader
- ONNX Runtime Go bindings
- Model optimization tools
- Cross-platform inference
TensorRT Bindings
PlannedNVIDIA TensorRT ecosystem bindings for high-performance inference on NVIDIA GPUs.
- TensorRT core bindings
- TensorRT-LLM for large language models
- TensorRT-RTX for consumer GPUs
- INT8/FP16 quantization support
Phase 3: Reinforcement Learning
Sugar
PlannedReinforcement learning framework for Go — environments, algorithms, and training utilities.
- Gym-compatible environments
- DQN, PPO, A2C algorithms
- Multi-agent support
- Distributed training
Want to contribute?
We welcome contributions! Check out our GitHub for open issues and discussions.
Join the Discussion