Spica on your screens
No screen of its own.
Spica shows up on the screens you already own.
Local AI for your home, out of the box.
In development

58 mm tall. Matte black, with a glass top and an amber light ring.
32 GB or 64 GB of LPDDR5X memory at up to 9600 MT/s, and a PCIe 5.0 SSD of 512 GB or 1 TB.
A fan draws it up across a vapor chamber and a radial heatsink.
58 mm tall. Matte black, with a glass top and an amber light ring.

32 GB or 64 GB of LPDDR5X memory at up to 9600 MT/s, and a PCIe 5.0 SSD of 512 GB or 1 TB.

A fan draws it up across a vapor chamber and a radial heatsink.

MoE, 20–40B
40–60tok/s
4 sessions of 1M tokens, 2 of 2M or 1 of 4M.
MoE, 40–100B
30–50tok/s
1 session of 128K–256K tokens, the longer for hybrid attention.
When the whole model fits in memory.
MoE, 100–300B
20–40tok/s
1 session of 1M tokens.
Experts streamed from the SSD at 12 GB/s, with a larger expert cache.
Dense or hybrid, 20–40B
20–40tok/s
1 session of 1M tokens, 2 of 512K or 4 of 256K.
Projected for the 64 GB Stone from the speeds we measure on the 32 GB laptop with its chip, with optimizations in progress (n-grams, radix attention). More than one session at once assumes continuous batching.
See today's resultsSpica on your screens
Spica shows up on the screens you already own.
What stays home
Models, chats and what it learns about you stay on the Stone, encrypted and working offline.
What's inside, filled in as each part is final.