# Into: Connito Most Bittensor subnets run inference on models that already exist. Connito is trying to do the harder thing: train frontier-scale language models from scratch, by splitting them into expert pieces and handing each piece to a different miner. // MoE models, trained by a swarm --- > New to Bittensor? Start here. Experienced users can skip to the full analysis. ### What is Connito? Connito, registered as Subnet 102, is a network for training large language models without a central data center. Instead of one operator renting a giant GPU cluster, the model is broken into specialized "expert" modules, and independent miners around the network each train a piece. Validators stitch the contributions together and score who did useful work. **The simple version:** Picture a giant brain that is too big to fit in any single computer. Connito chops the brain into specialist sections, one good at code, one at math, one at language, and lets different people train each section on their own hardware. A router learns which specialist to ask for any given question. **Centralized equivalent:** The frontier labs (OpenAI, Google, Meta) that train huge models on tightly controlled GPU clusters. Connito is attempting the decentralized version of that, where no single party owns the compute. **How it works:** - **Miners** train expert modules of a Mixture-of-Experts model on their own hardware and submit their contributions, per the subnet's repository. - **Validators** coordinate the training rounds, aggregate the submitted contributions, and score miners. Those scores set the weights that drive how emissions are distributed. --- ### Why This Matters - **The problem it solves:** Training a 100-billion-parameter model is out of reach for almost everyone, because it needs a cluster few can afford. Connito's bet is that a frontier-scale Mixture-of-Experts model can be partitioned into pieces small enough for individual miners to fit and train. - **The opportunity:** If decentralized training actually works at scale, it changes who gets to build large models: not just the handful of labs with the biggest GPU budgets, but anyone who can contribute an expert module. - **The Bittensor advantage:** Mixture-of-Experts is a natural fit for a subnet. The architecture is already modular, so the work splits cleanly across independent miners, and Bittensor's incentive layer pays for the contributions that measurably improve the model. - **Traction signals:** Connito published a whitepaper and shipped what it calls a live alpha code release. The public repository shows active development, and the subnet has drawn discussion from Bittensor commentators. It is early: the active miner set is small and the model is still in its first months. --- ## Full Analysis **Category:** Distributed Training | **Centralized Competitor:** Frontier AI labs (OpenAI, Google, Meta), and within Bittensor, Macrocosmos (SN9) Decentralized model training is one of the oldest ambitions on Bittensor and one of the hardest to pull off. Coordinating many untrusted machines to train one coherent model, then verifying that each contribution actually helped, is a genuinely difficult problem. Connito's angle is to lean on a Mixture-of-Experts design, where the model is already a collection of separable specialists, and turn open-source expert updates into something the network can reward. **Mechanism:** Per the subnet's repository, Subnet 102 splits a large language model into expert groups distributed across many independent miners, rather than training one monolithic model on a single machine. Miners train their assigned expert modules; validators coordinate the rounds, aggregate the contributions, and score miners. Those validator scores set the weights that determine how the subnet's alpha emissions flow. Connito's public materials describe the goal as expert partitioning that lets a 100-billion-parameter-plus model be trained in pieces an individual miner can actually fit, with a router learning which expert to call. The development picture is consistent with a team in active build mode. The public GitHub repository was created in early April 2026 and last pushed on 9 June 2026, with recent commits touching validator memory handling, a model backend port, and telemetry for a leaderboard dashboard. The work comes from a small group of contributors, led on commit volume by the account isabella618033, who is also listed as the subnet's on-chain contact. The codebase is primarily notebooks and Python, which is typical for a training-focused subnet. On the market side, Connito's alpha token is young and its price is still finding a level: it is down sharply over its first month, and net staking flows have been soft over the past week across the data sources we checked. The pool holds a few thousand TAO in depth. Root proportion sits near 0.6, which is what you would expect from a subnet only a couple of months old whose price has not settled. These are early readings, not a verdict. --- ### Risk Factors - **Execution risk:** Decentralized training at frontier scale is unproven. The 100-billion-parameter ambition is a claim about where Connito is headed, not a demonstrated result, and coordinating expert training across untrusted miners is one of the harder things to get right on Bittensor. - **Early stage and market:** The token is roughly two months old, the active miner set is small, and the price has fallen notably over its first month. Early subnets are volatile, and current emission share is small. - **Deregistration:** Connito registered on 28 March 2026, so it sits inside the four-month immunity window (until roughly late July 2026) and cannot be deregistered yet. After that, sustaining positive net staking flows becomes the test that keeps the slot. - **Competition:** Distributed training is a crowded lane on Bittensor. Connito's own materials position it against Macrocosmos (SN9), and it will be measured against established training subnets. --- Into the next one.