About 20 flipped bits are enough to break a language model

A language model holds billions of parameters. Corrupting roughly twenty of the bits inside them, at random, is enough to stop it working.

What happened

Benedikt Holm published the experiment on August 20, 2026. The method was direct: flip random bits in a model’s weights, one at a time, and measure when the output stops being coherent. The primary subject was Qwen2.5-Coder-3B, with the work repeated on Mistral-7B, Qwen3-8B, Granite-4-8B and phi-4.

In FP16, the models broke after a median of about 23 random flips. The failure mode was consistent across runs: endless token loops in which the model repeats the same characters indefinitely, early emission of a stop token, and incoherent repeated strings.

The distribution of risk turned out to be extremely concentrated. Bit 14 in FP16 is the most significant bit of the exponent, and a single flip there can change a weight from 0.021 to 1352, which is enough to destabilize the network on its own. When Holm shielded that one bit and left the other fifteen unprotected, the same models tolerated between 79,000 and 490,000 flips before failing.

Quantization changed the picture as well. A Q4_K_M build required a median of 1024 flips against 22 for FP16, approximately 49 times more resilient.

What it means

Radiation-induced bit flips are a real engineering constraint in low Earth orbit and at altitude, where commercial memory is exposed to particle strikes. They are also relevant on the ground for consumer hardware running local inference over long periods without error-correcting memory.

The practical value of the result is the concentration. Protecting one bit position out of sixteen removes most of the exposure, which makes selective hardening cheap compared to full ECC. The secondary finding is the more interesting one for anyone running models locally. Quantization was adopted to reduce memory footprint, not to improve fault tolerance, and it turns out to deliver a large robustness gain as a side effect of using fewer and differently distributed exponent bits.