Summary
According to a post on X by @bnjmn_marie, Benjamin Marie of The Kaitchup ran ISTA's Statistically-Lossless Quantization search on an RTX 3060 for 38 hours. The search mixed tensors from Liquid AI's and AtomicChat's GGUF files for Liquid's LFM3 1.7B. The resulting model scores 97.74% on EAR@10, a measure of how closely it matches the full model, versus 97.52% for AtomicChat's AD-Q6_K, both at 1.588 GB. Marie says the real-world difference may be small.
AI-assisted summary based on the listed source.
What happened
Benjamin Marie (The Kaitchup) ran ISTA's Statistically-Lossless Quantization search on an RTX 3060 for 38 hours, mixing tensors from Liquid AI's and AtomicChat's GGUF files for Liquid's LFM3 1.7B. The result scores 97.74% EAR@10, a measure of how closely it matches the full model, against 97.52% for AtomicChat's...
Why it matters
For people running local LLM inference on consumer GPUs, this shows that automated search over quantized tensor mixes can shift benchmark scores at a fixed file size. The gap is 0.22 percentage points, and the author himself suggests it may not show up in practice. Readers should watch for independent evaluations before treating the model as a clear upgrade.
Signal Intelligence
Signal Strength 78%
Technical label WATCH
Public Interest 25
Category DEVELOPING
Reader Depth TECHNICAL
Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.
Public Interest components
Recognizable Entity Score 0
Practical Impact Score 18
Novelty Interest Score 70
Consequence Score 10
Curiosity Score 16
Shareability Score 44
Why this is here
VQV surfaced this signal because it is recent, relevant to LLM Inference, connected to Scobleizer News Pulse (X).
