Subnivean
ManifestoResearch logTalk to us

Research log

Research log

Every entry, including the dead ends.

No.

Date

Entry

Field

Length

01

14 Oct 2026

Stacked scanline layers, grey above and colour below

Clever quants for more work per watt

Mixed-precision ternary quantization of GLM-5.3 Flash: about 10% less memory for about 3% quality.

Research

8 min

Subnivean

An inference engineering lab building a vertically integrated serving stack for open‑weight models.

Site

Home

Manifesto

Research log

Talk to us

Compliance

Terms

Privacy

© 2026 Subnivean

hello@subnivean.xyz