Russian startup Mostik (“little bridge”) has taught different AI models to exchange not text, but weight values — parameters that determine how a prompt is turned into a response. WIRED author Will Knight calls this approach “machine telepathy.” Using it, the company has already built a model that reached first place in the complex ARC-AGI 3 competition, and in a demo connected the open GLM-5.2 with 753 billion parameters to the mobile Qwen-3.5 with 4 billion parameters, so that the hybrid costs 20 times less than the full large model. The startup is not disclosing technical details until the end of the competition.
What happened
The approach is described in a piece by Will Knight in the WIRED AI Lab newsletter. The mechanism is described as follows: models exchange weight values with each other, so the large model no longer needs to process the entire inference cycle — its quality is passed directly to the small model through the weights. In the demo, Mostik connected the largest version of the Chinese open GLM-5.2 with 753 billion parameters to a 4-billion Qwen-3.5 running on mobile devices; according to the company, the quality of the combination is exactly in the middle between the two models at a price 20 times lower than the full GLM-5.2. Separately, using this approach, the startup built a model that reached first place in the ARC-AGI 3 competition: the team is deliberately not disclosing technical details in order to win the competition, and promises to share details after it ends.
Context
Weight exchange is a separate class of model communication, different from familiar ensembles and distillation, although it is currently impossible to assess the scientific novelty of the approach without a preprint and algorithm description. The company is backed by recognizable names: CEO Sasha Malysheva developed the approach herself, and the position of chief scientific officer is held by Stanislav Smirnov — a 2010 Fields Medal laureate and professor at the University of Geneva. The startup’s name — “little bridge” — directly refers to its main idea: a bridge between models. ARC-AGI 3 is a complex competition for AI systems, and as of today the approach has essentially two public pieces of evidence: the competition result and the WIRED piece.
Why this matters for the industry
If “bridges” between models scale, progress may come not from monolithic giants, but from networks of specialized models — Sasha Malysheva directly says that she does not believe in the future of a single monolithic model thanks to scaling. For builders, this is a signal for the architecture of “local small model plus remote quality supply from a large one”: if the approach is at least partially reproduced and becomes available, mobile and edge products will be able to offer mid-level quality with the savings against calling a flagship model that is claimed in the Mostik demo. The value of open weights and narrow-domain models, for example in biology and physics, also increases, in contrast to the closed solutions of Anthropic and OpenAI. However, for now, the market has neither an API nor pricing: this is a statement about a direction, not an available tool.
Why this matters for users
The only verifiable action for a reader now is to watch the ARC-AGI 3 leaderboard: that is where the first independent data on the approach should be expected. The demo combination itself suggests that strong quality can work even on a phone: the 4-billion Qwen-3.5 receives “supply” from a giant model, which points to cheap local hybrids instead of expensive API subscriptions. It is reasonable to already recalculate one’s own scenarios of “expensive quality via API versus cheap local hybrid” and prepare an eval set on which such hybrids can be tested. But today this is a demo and an article, not a product: no access, prices, or deadlines have been announced.
What is still unknown / limitations
All key demo figures are provided by the company itself and currently exist only in a single WIRED piece: the methodology is hidden until the end of ARC-AGI 3, there is no preprint, code, latency data, or independent reproductions, and third-party confirmations are limited to familiarity with the technology. The smooth formulation of quality “exactly in the middle” without a measurement protocol is indistinguishable from marketing, and the demo itself is compatible with ordinary distillation, so the scientific novelty of weight exchange remains unproven. Malysheva’s statement about the end of the era of a single monolithic model is a hypothesis about the industry trajectory that one demo does not prove. Until details are disclosed and independently verified, this is a claim, not an established result.
Sources
Author
Look at AI, editorial team