VQV Signal

ROBOTS & HARDWARE WATCH TECHNICAL

Thinking Machines reports training Inkling 2 Small with a native 4-bit forward pass

According to a post on X by @thinkymachines, a Connectionism post by Jack Cook and colleagues describes using NVFP4 for forward passes in mixture-of-experts layers, and MXFP8 with re-quantization and stochastic rounding for backward passes, along with quantized GPU communications. The lab says Inkling 2 Small is the largest model it knows of trained end-to-end this way. It reports wall-clock speedups of 8% in pretraining and 20% in RL, though the excerpt is cut off before the comparison baseline is stated.

Source: Scobleizer News Pulse (X) · x.com Found via Robert Scoble’s Scobleizer News Pulse · a post on X, not independently verified Published 2026-11-15T20:45:00+00:00 Detected 2026-11-15T22:24:04+00:00
View original source

According to a post on X by @thinkymachines, a Connectionism post by Jack Cook and colleagues describes using NVFP4 for forward passes in mixture-of-experts layers, and MXFP8 with re-quantization and stochastic rounding for backward passes, along with quantized GPU communications. The lab says Inkling 2 Small is the largest model it knows of trained end-to-end this way. It reports wall-clock speedups of 8% in pretraining and 20% in RL, though the excerpt is cut off before the comparison baseline is stated.

AI-assisted summary based on the listed source.

A Connectionism post by Jack Cook and colleagues at Thinking Machines explains how they used NVFP4 in the mixture-of-experts layers for forward passes, MXFP8 with re-quantization and stochastic rounding for backward passes, and quantized GPU communications. The lab says Inkling 2 Small is the largest model it...

Teams training large mixture-of-experts models could be affected, because the post describes lower-precision arithmetic across pretraining and RL, not only at inference. The speedup figures come from the lab's own post, so anyone weighing training costs should check the full write-up for the baseline and whether other groups reproduce the results.

Hardware and robotics watchers may want to track whether this becomes a product, benchmark, or deployment signal.

Signal Strength 81% Technical label WATCH Public Interest 18 Category ROBOTS & HARDWARE 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 0 Novelty Interest Score 70 Consequence Score 10 Curiosity Score 0 Shareability Score 21

VQV surfaced this signal because it is recent, relevant to LLM Inference, connected to Scobleizer News Pulse (X).