enjoy-digital/trixium is an open source project that decodes PrismML's Bonsai-family ternary LLMs on an FPGA card. The repository description says the host handles tokenization while model decoding runs inside the card. There is no public release version. Weights are expressed as -1·0·+1, so the computation uses addition, skipping and subtraction instead of ordinary multiplication.
The project provides an end-to-end implementation that reaches model decoding inside the FPGA card, and treats LiteX, LitePCIe, LiteDRAM and VexiiRiscv as its own components. The supported targets are Ternary-Bonsai 1.7B and 8B, and Ternary-Bonsai-2 2B, 4B and 27B. The repository description introduces the first two models as Qwen3-based and the rest as a Qwen3.5 hybrid that combines DeltaNet with attention. The design centers on memory bandwidth and is aimed at single-user chat. Even without a card, the whole SoC can be run in Verilator, so the execution path can be examined without hardware. The license is BSD-2-Clause and the primary language is Python.
The repository publishes measured processing speeds per card, but those measurements are based on three card types. It is therefore hard to generalize the results to the performance of other FPGA cards. This material is better read as a basis for judging the implementation in a specific card configuration.