Meet Baltoro
The RISC-V First AI Compiler Stack
Baltoro is a full-stack AI compiler solution purpose-built for RISC-V Vector processors, custom RISC-V extensions, and domain-specific accelerators.
Built on MLIR at its core, Baltoro combines maximum performance with the flexibility needed to keep pace with rapidly evolving AI workloads.
Its modular architecture — Frontend, Compiler, and Runtime — provides a seamless path from high-level ML frameworks to highly optimized inference across heterogeneous RISC-V systems. When combined with QuantX, Baltoro delivers compressed models allowing developers to unlock the full potential of RISC-V for AI.
What you need to know about QuantX?
QuantX streamlines the path from research to deployment by making it simple to prototype new ideas, evaluate performance, and unlock the full potential of compressed models. Built on both state-of-the-art research and our own innovations, QuantX delivers models finely tuned to your hardware.
Variants
Baltoro
- Out-of-box compilation of major vision and generative AI models on standard RISC-V scalar and vector
- Bring your own model and we enable it in Baltoro
Baltoro+
- Baltoro+ enables all upcoming AI extensions to the RISC-V standard like P, IME and AME
- We also enable program lowering using your custom AI extensions in Baltoro+
Baltoro++
- We enable your neural processing unit (NPU) with customized MLIR lowering and code generation in Baltoro++.
QuantX: Scalable AI Optimization for Peak Performance and Efficiency
We support mixed-precision and multi-granularity quantization enhanced with outlier reduction and non-uniform sampling, advanced numeric types such as micro-scaling and block floating point, and structured pruning — with more optimizations on the way.
Designed for hardware efficiency, QuantX works with minimal changes to existing model graphs, and its modular design makes it easy to adapt to custom needs or integrate with proprietary SDKs.
Backed by our team’s deep expertise and cutting-edge optimization strategies, QuantX empowers you to achieve performance and efficiency at scale — finding the best path forward for your models and hardware.
Latest Updates
LLM Inference with Codebook-based Q4X Quantization using the Llama.cpp Framework on RISC-V Vector CPUs
We utilize QuantX [MN25], our in-house hardware-aware quantization platform, to generate a