AI/ML●●●Banger
Glq LLM quantization using E8 lattice
E8 lattice codebooks beat GPTQ at 2-4 bpw with fused CUDA kernel skipping weight materialization.
WizardryBig Brain
acd
2013d ago
Back to browse
3.125-Bit LLM quantization bypassing tensor cores
by dmaniss·May 21, 2026·3 points·0 comments
AI Analysis
●●●BangerBig BrainBold Bet
Replaces Tensor Cores with LUTs and bitwise ops for 3-bit edge inference.
Strengths
- •Targets memory bandwidth bottleneck instead of compute, which is the real constraint at batch size 1.
- •Data-free quantization avoids needing calibration datasets, simplifying deployment.
- •Challenges the industry assumption that more Tensor Cores are the only path forward.
Weaknesses
- •No code, benchmarks, or proof-of-concept implementation provided yet.
- •Claims about preserving reasoning capabilities need empirical validation.
Category
Target Audience
ML researchers, embedded systems engineers, AI hardware architects
Similar To
AWQ · GPTQ · GGUF
Similar Projects
AI/ML●●●Banger
1-Bit Bonsai, the First Commercially Viable 1-Bit LLMs
1-bit weights matching 8B model performance while running 132 tokens/sec on M4 Pro.
Big BrainZero to OneWizardry
PrismML
4301532mo ago
AI/ML●●●Banger
Running a 1.7B parameters LLM on an Apple Watch
Runs a 1.7B LLM offline on Apple Watch using 1-bit quantization.
WizardryNiche Gem
pielouNW
302mo ago
AI/ML●●●Banger
Core Rth. A governed AI kernel for engineers who don't trust their LLMs
Proposal-first governance + hardware E-stop for AI controlling robots/drones—legitimately novel safety architecture.
Big BrainBold Bet
christianrth
113mo ago
AI/ML●Mid
The Deterministic Core Architecture for AI-Augmented Applications
Interesting diagnosis of AI statelessness, but six artifacts aren't directly accessible.
Big Brain
Brandon_Bell
309d ago
AI/ML●Mid
Fixing AI's Core Flaws, A protocol cuts LLM token waste by 40–70%
The repo outlines a concrete seven-layer protocol (SLP, World Model Interpreter, Agent/Persona/Knowledge layers, Metacognition, etc.) and even splits each piece into its own subrepo — that modular breakdown is the repo's strongest move. But this reads more like an ambitious manifesto and design spec than a working system: good docs and diagrams are present, yet there's little visible implementation, benchmarks, or reproducible evidence for the bold claims (like 40–70% token savings).
Bold BetRabbit Hole
WujieGuGavin
103mo ago