Technology
Startup Mostik develops method for AI models to communicate through weight values
The direct mathematical approach aims to bridge large and small AI systems efficiently without generating intermediate text outputs.
The short version
- Startup Mostik created a technique allowing artificial intelligence systems to interact directly using internal weight values rather than standard text generations.
- A demonstration combining a 753-billion-parameter model with a 4-billion-parameter model operated at one-twentieth the cost while achieving intermediate performance.
- The startup has kept specific details of its architecture confidential while competing in the ARC-AGI 3 contest, leaving its broader viability and generalizability uncertain.
Key facts
- Mostik was founded by CEO Sasha Malysheva, with 2010 Fields Medalist Stanislav Smirnov serving as chief scientist.[Wired]
- The company's method enables AI models to transfer capabilities without having to convert outputs into text tokens first.[Wired]
- In a demonstration bridging a 753-billion-parameter GLM-5.2 model and a 4-billion-parameter Qwen-3.5 model, the combined system operated at roughly 5% of the large model's cost while achieving mid-level performance.[Wired]
- Mostik has entered a model using this technique into the ARC-AGI 3 benchmark competition.[Wired]
What remains uncertain
- Full technical specifications of Mostik's implementation have not been made public due to active participation in the ARC-AGI 3 competition.[Wired]
- Whether the technique can scale broadly across diverse, closed, or proprietary model architectures remains unverified outside initial demonstrations.[Wired]