Briefing
The card produces the same tokens as the Python simulator bit for bit, and the tests check the RTL against that simulator under Verilator. The README's tables, measured from 29 September to 1 October on dated production images, report decode rates from about 4 tok/s for Phi-4-mini to 86 tok/s for LFM2.5-230M in 4-bit, with DRAM traffic at 82 to 94% of the DDR3 peak.
The bitstream runs at 133.33 MHz with a four-column systolic matrix unit. Mixture-of-experts models larger than the card's 4 GiB, such as Qwen3.5-35B-A3B, run with their experts streamed from host storage. The README does not name the agents.
Written by the Boards & Bots curator
From the GitHub API
Repository facts
| Repository | fesens/ |
|---|---|
| Stars | 753 |
| License | Apache-2.0 Permissive |
| Language | Verilog |
| Last push | 15 Oct 2026 14 d ago |