Scaife, BSHJ and Culet
The parts of Lamproite, our TPU software platform, that run under the Lamproite API and every Lamproite Pack.
Scaife
Our engine: the runtime that serves our API and every Pack.
- On Llama 3.2 1B with the same weights: 47,895 tokens a second at 256 conversations, 2.10 times Google's vLLM-TPU 0.31 (22,826). Measured 19 October 2026.
- EdenLive 5B by VAST (built on Wan2.2) runs live at 1280x704 and 39.4 frames a second in our engine, with the first chunk in 256 ms. EdenLive's own report gives 29.8 frames a second on an NVIDIA H100. Measured 22 October 2026.
- Xiaomi MiMo-V2.6-Flash: 180.4 tokens a second for one conversation, 1.22 times the fastest provider listed on Artificial Analysis (Makora, 147.8), with a 10,000-token prompt and 5.31 s to first token. Measured 22 October 2026.
BSHJ
Our weight format: the container for model weights, loaded directly by Scaife. Load-time results on your own model are shared under NDA.
Culet
The low-level code under Scaife.
- Llama 2 7B at 4-bit through llama.cpp on TPU: 249.9 tokens a second, above the RTX 4090 (188.96) and A100 80GB (200.90) rows of llama.cpp's scoreboard. Measured 19 October 2026.
- Our llama.cpp backend for TPU runs the GGUF files you already have. In our coverage run, 44 of 47 GGUF model files passed our correctness check against llama.cpp's own CPU backend. Benchmark: our coverage run, 19 October 2026.
- Five merged pull requests in llama.cpp, and two fixes sent upstream to Google's MaxDiffusion.
Each result is copied word for word from our research page or our home page, with its date.