GoKawiilStartup Mostik lets AI models share intelligence directly via weight values, not text
Russian startup Mostik has developed a method allowing AI models to exchange information through their internal weight values rather than generated text, effectively letting a smaller model absorb capabilities from a larger one. The team demonstrated this by linking a 753-billion-parameter GLM-5.2 model with a 4-billion-parameter Qwen-3.5 model, producing a hybrid system that runs at one-twentieth the cost of the full-size model while performing roughly midway between the two in capability. Mostik also used a related technique to build a model that has topped the ARC-AGI 3 benchmark, though details remain undisclosed while the contest is ongoing.
GoKawiil's interpretation of the reporting above, not reported fact.
Most multi-model AI systems today combine outputs by feeding one model's text into another, which is slow and expensive. A direct, non-verbal exchange of information between models could make it far cheaper to boost smaller models' performance without running massive ones, potentially reshaping how companies deploy AI at scale and on limited hardware like phones.
- Mostik's method lets AI models share capabilities via their weight values instead of text output.
- A hybrid combining a 753B and a 4B parameter model cost 1/20th of the larger model with mid-range performance.
- The technique also powers a model currently leading the ARC-AGI 3 benchmark, details undisclosed.
NVIDIA GeForce RTX 4090 Founders Edition — If reading about bridging huge open-weight models to tiny ones makes you want to run your own, a RTX 4090 is the go-to card for local LLM experimentation. Its large VRAM pool lets you load quantized open models like Qwen and tinker with weights yourself instead of just reading about it.
See NVIDIA GeForce RTX 4090 Founders Edition on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.Source: wired.com — Will Knight, 2026-09-02
Published there as: “These Russian Mathematicians Taught AI Models How to Talk to Each Other Without Using Words”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.