Mostik's stealth-exit writeup on passing hidden states between frozen models through a trained bridge, so a 4B model decodes with a 753B model's latent context instead of its text.
Mostik, built by friends of mine, came out of stealth on 2 September with this writeup. One model reads the problem and stops without writing a word; its hidden states cross a small trained bridge into a second model, which does all the decoding. No text passes between them and both models stay frozen, so the bridge is the only part of the system that learns anything.
The reason to bother is what a model throws away. To produce one token it builds about two megabytes of internal state, then picks one entry from a vocabulary of around 150,000 and discards the rest. Seventeen bits leave. Every council, router and coding subagent you've ever wired together runs on those seventeen bits.
The first result pairs GLM-5.2 (753B) as sender with Qwen-3.5 (4B) as receiver, an expensive prefill and a cheap decode. Mostik reports that the bridge closes 50% of the distance between the small model's score and the large one's, and that the bridged pair costs 2.5x less compute than the mid-sized model which would score the same.
It's all self-reported, there's no paper linked, and the writeup never names the benchmark, so "50% of the gap" is carrying weight nobody outside the company can check. WIRED has the first external account.