LOCAL AI · AGENT & INFERENCE RESEARCH
Intelligence you own,
not rent.
We’re building advanced local AI to help you do more with the hardware you already have.
Iris on your computer
New chatIris · Preview
Iris
Reading papers
Selecting relevant passages
＋Ask anything
Four papers, one request.
Iris selects relevant passages and reads in parallel.
Animated interface concepts · recordings coming soon
ENGINE PERFORMANCE · OCTOBER 2026 SNAPSHOT
The numbers.
20–49%
higher aggregate throughput
vs Ollama 0.40.1
RTX 5070 Ti · 8–16 concurrent chats
17 October 2026
17 October 2026
114ms
to the first token
vs 160 ms with llama.cpp
Apple M5, 24 GB · 52-token prompt
9 October 2026
9 October 2026
Across workloads and test conditions
RTX 5070 Ti · 8–16 chats · throughput vs llama.cpp+12.7–15.1%
RTX 5070 Ti · one chat · first token vs Ollama2–2.5× slower
Apple M5 · 4 chats · throughput vs llama.cpp−3.8%
Figures transcribed from the October ledger; different machines and workloads.
