Apex
SN1Compete to build the best AI algorithms, with rewards for the winners
The original Bittensor subnet, now reinventing itself as a competitive arena where algorithmic intelligence is the product.
// Where code competes and the best algorithm wins
Apex (SN1) is Bittensor's longest-running subnet, operated by Macrocosmos. It hosts structured competitions where miners submit Python-based algorithms to solve specific problems: currently reinforcement learning (Battleship) and distributed training optimization (IOTA Simulator). Validators evaluate submissions in sandboxed environments and reward the top-performing solutions with emissions.
The simple version: It's like Kaggle, but on-chain. Submit code, get scored, and the winner earns alpha emissions.
Centralized equivalent: Kaggle competitions, but with continuous payouts and direct integration into Bittensor's distributed training stack.
How it works:
- Miners submit algorithmic solutions (Python code or model weights) to active competitions via a CLI
- Validators run submissions in sandboxed environments, score them against objective metrics, and assign emissions to the top performer
- The problem it solves: Building optimal algorithms for distributed systems is expensive and slow. Apex crowdsources algorithmic R&D by incentivizing global talent to compete on well-defined problems, producing production-ready solutions faster than traditional internal development.
- The opportunity: The IOTA Simulator competition is a direct pipeline into real distributed training infrastructure. If Apex can consistently produce algorithms that improve IOTA's training speed and efficiency, it becomes the R&D arm for one of Bittensor's most ambitious subnets. The competition framework itself is being designed toward self-serve creation, which could open it to external customers.
- The Bittensor advantage: Permissionless global participation means anyone can submit a solution. The emission incentive attracts talent that wouldn't otherwise contribute to open-source distributed training research. The winner-takes-all model creates strong competitive pressure.
- Traction signals: The Matrix Compression competition delivered measurable production value (~3x compression improvement integrated into IOTA). The IOTA Simulator launched with immediate community engagement. GitHub shows 2,867 commits across 29 contributors with recent activity (last commit March 26, 2026). However, the winner-takes-all model means only 3 hotkeys earn incentive at any given time, which limits the breadth of active participation.
Category: Reinforcement Learning and Subnet Composition | Centralized Competitor: Kaggle, HuggingFace competitions, CrunchDAO
Apex occupies a unique position in the Bittensor ecosystem. As SN1, it carries historical weight, but its current identity is defined by a pivot toward structured algorithmic competitions that feed directly into Macrocosmos' broader product stack. The subnet isn't just running competitions for competition's sake: its Matrix Compression series produced a ~3x reduction in data transmission size that now accelerates IOTA's distributed training pipeline.
Mechanism:
Apex runs multiple competitions in parallel, each with its own emission allocation. The model is winner-takes-all: only the top scorer in each competition earns emissions, and new submissions must beat the current leader by at least 1% to claim the top spot. This prevents trivial code copying once solutions are revealed at round end. Competitions run on fixed round lengths (1 day for the IOTA Simulator, longer for Battleship), and submissions are evaluated in CPU-only sandboxed containers with strict memory and time limits. No GPU is required to compete.
The current competition lineup includes two active tracks. The RL Battleship competition has miners training reinforcement learning agents to play Battleship, with recent enforcement against heuristic-only submissions disguised as neural networks. The IOTA Simulator competition, launched in late March 2026, is the more consequential track: miners submit routing and balancing algorithms that orchestrate activations across a simulated network of heterogeneous nodes, scored on how efficiently they move forward and backward passes through layered pipelines under realistic conditions including node churn, variable latency, and bandwidth constraints. This competition directly serves IOTA (SN2), Macrocosmos' distributed training subnet, making Apex a research and development engine for the broader Macrocosmos stack.
The team plans to roll out new competitions on roughly a monthly cadence, with quality prioritized over quantity. The Matrix Compression competition was recently sunset after its winning algorithm was integrated into IOTA's production pipeline.
- Winner-takes-all concentration: Only the top scorer per competition earns emissions. This creates a high barrier for new entrants and means the vast majority of miners earn nothing. Community members have flagged this as discouraging, particularly when a top score appears to have plateaued.
- Closed-source validator pipeline: The evaluation infrastructure is proprietary. Community members have raised concerns about this lack of transparency, and it limits independent verification of scoring fairness. The team has acknowledged this tradeoff but has not indicated plans to open-source the pipeline.
- Infrastructure reliability: Discord logs reveal recurring sandbox issues: CPU throttling, dropped jobs, evaluation delays spanning hours, and bugs in new round data (e.g., invalid bfloat16 buffers in Matrix Compression). The team is responsive to bug reports but the pattern suggests the evaluation infrastructure is still maturing.
- Low emission share and liquidity: At 1.22% emission share, a momentum score of 12.5/100, and relatively thin volume (2.7K TAO daily against a 55.9K TAO market cap), Apex is not commanding significant capital flows despite its historical significance as SN1.
- Ecosystem dependency: Apex's highest-value competitions are tightly coupled to IOTA and Macrocosmos' internal needs. If IOTA's direction shifts or Macrocosmos deprioritizes Apex, the competition pipeline could stall.
