TT - Forge - 2026 Pricing, Features, Reviews & Alternatives

TT - Forge is an open-source, MLIR-based end-to-end compiler stack that bridges high-level machine learning frameworks with Tenstorrent AI accelerator hardware. Currently in public beta, it compiles models from PyTorch, JAX, TensorFlow, and ONNX, lowering them through optimized intermediate representations into executable workloads via TT-NN and TT-Metalium. Deployment options include on-premises installation via pip wheel or Docker container on Tenstorrent-equipped systems, cloud deployment via Koyeb, and source access through GitHub.

The stack is composed of four primary components: TT-XLA, a PJRT-based bridge supporting JAX and PyTorch/XLA with multi-chip execution; TT-Forge-ONNX, a framework-agnostic frontend powered by TT-TVM for ingesting ONNX, TensorFlow, and similar frameworks; TT-MLIR, the middle and backend compiler featuring custom dialects (TTIR, TTNN, and TTKernel) for hardware-aware optimization targeting high utilization and efficient memory access; and TT-Blacksmith, a cookbook for LLM fine-tuning and training experiments on Tenstorrent devices. An automated compilation path covers a broad range of model architectures, including LLMs and CNNs, without requiring custom kernel development.

TT - Forge targets ML engineers, AI researchers, and software developers who need to run deep learning workloads on Tenstorrent silicon using standard frameworks. The modular, open-source architecture integrates with the broader OpenXLA, MLIR, ONNX, TVM, PyTorch, and TensorFlow ecosystems, supporting adoption of new ops, frameworks, and hardware targets as those ecosystems evolve. TT-Explorer, a visual performance analyzer, provides interactive inspection of model graphs, memory plots, tensor and buffer views, and execution traces to support model optimization workflows.