VQV Signal

ROBOTS & HARDWARE WATCH TECHNICAL

Thinking Machines details end-to-end 4-bit training recipe for Inkling 2 Small

According to a post on X by @iamgrigorev, Thinking Machines has published a write-up titled "End-to-End 4-Bit Training," by Jack Cook and colleagues. It describes how Inkling 2 Small (594B total, 17B active) was trained with a native NVFP4 forward pass, from the first pretraining step through the final RL rollout. The approach uses NVFP4 in the expert layers, MXFP8 backward passes with re-quantization and stochastic rounding, and quantized GPU communication. The lab reportedly cites an 8% figure, but the available text is cut off before it explains what that number measures.

Source: Scobleizer News Pulse (X) · x.com Found via Robert Scoble’s Scobleizer News Pulse · a post on X, not independently verified Published 2026-11-13T23:45:00+00:00 Detected 2026-11-14T06:25:22+00:00
View original source

According to a post on X by @iamgrigorev, Thinking Machines has published a write-up titled "End-to-End 4-Bit Training," by Jack Cook and colleagues. It describes how Inkling 2 Small (594B total, 17B active) was trained with a native NVFP4 forward pass, from the first pretraining step through the final RL rollout. The approach uses NVFP4 in the expert layers, MXFP8 backward passes with re-quantization and stochastic rounding, and quantized GPU communication. The lab reportedly cites an 8% figure, but the available text is cut off before it explains what that number measures.

AI-assisted summary based on the listed source.

Thinking Machines posted "End-to-End 4-Bit Training," by Jack Cook and colleagues, explaining how Inkling 2 Small (594B total, 17B active) was trained with a native NVFP4 forward pass from the first pretraining step through the final RL rollout: NVFP4 in the expert layers, MXFP8 backward passes with...

Teams training large models on low-precision hardware could use this description as a reference for applying 4-bit formats across pretraining, backward passes and reinforcement learning. The snippet only partly shows the write-up, so readers should check the original post before relying on the 8% result or its basis. Watch for independent testing or replication of the reported results.

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

Signal Strength 83% Technical label WATCH Public Interest 39 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 51 Practical Impact Score 0 Novelty Interest Score 94 Consequence Score 10 Curiosity Score 0 Shareability Score 56

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