// tag
#llm
2026
TriXium, part 3: from Bonsai to Bonsai 2, porting an FPGA LLM design to a hybrid architecture
Part 3 of TriXium: porting a design built for the Qwen3-based Bonsai models to the Qwen3.5 hybrid Bonsai 2 family. The CPU path first, token-exact in …
TriXium, part 2: a language model on an Efinix Ti375, from power-up to chat in 20 seconds
Part 2 of TriXium: moving from PCIe cards to the small Efinix Ti375 kit. LPDDR4x at its full rate through both AXI targets, the weights over UDP, then …
TriXium, part 1: ternary LLMs on second-hand FPGA cards, as fast as the memory allows
Part 1 of TriXium: ternary LLMs decoded on FPGA cards. Large cards first (FK33 HBM2, XCU1525 4x DDR4) to get everything in place and memory-bound, …
blogcontact