Apex
SN1Compete to build the best AI algorithms, with rewards for the winners
Bittensor's oldest subnet is now a competition arena where Python algorithms fight for emissions, with RL Battleship and a distributed training simulator running side by side.
// Where code competes and the best algorithm wins
Apex (SN1) is the original Bittensor subnet, operated by Macrocosmos. It runs structured Competitions made up of multiple Rounds, where miners submit Python algorithms through the Apex CLI and validators score them against benchmark tasks. The current competitions, per the repo, are RL Battleship (a reinforcement learning Battleship game on a 10x10 grid) and Iota Simulator (a distributed training simulation where miners write routing and balancing algorithms for heterogeneous nodes).
The simple version: Think Kaggle, but the leaderboard pays out in alpha tokens and the problems are rotating, on-chain Bittensor challenges.
Centralized equivalent: Kaggle competitions, with the added twist that scoring runs continuously on-chain and the prize is subnet emissions instead of cash.
How it works:
- Miners submit Python algorithms via the Apex CLI to the active competitions (currently RL Battleship and Iota Simulator) and run them on the subnet.
- Validators evaluate each submission against the competition's benchmarks, score them, and publish on-chain weights that drive emissions to the strongest entries.
- The problem it solves: Open algorithmic research has no shared, trust-minimised scoreboard. Apex makes the scoreboard itself the reward mechanism, with on-chain payouts replacing prize pools.
- The opportunity: The two live competitions point at two of the hardest open problems in AI infrastructure: efficient reinforcement learning agents and pipeline-parallel training across unreliable nodes. Solutions here generalise far beyond the games used to score them.
- The Bittensor advantage: Continuous, code-based competition with on-chain settlement gives a tighter feedback loop than periodic centralised contests, and rewards persist as alpha rather than one-off cheques.
- Traction signals: Apex is shipping. The repo released v4.1.32 on 2026-05-15 and has been pushed as recently as 2026-05-20, with 22 contributors and 135 stars. Macrocosmos runs four other subnets in addition to this one.
Category: Reinforcement Learning | Centralized Competitor: Kaggle, AIcrowd
Apex is Bittensor's longest-running subnet, registered as netuid 1 at the network's launch. Macrocosmos rebuilt it into a competition framework: rather than serving one fixed workload, the subnet rotates problem domains, with miners staking their submissions against rolling benchmarks and validators arbitrating scores. The current rotation, RL Battleship and Iota Simulator, picks at the seam between agentic decision-making and distributed-systems engineering.
Mechanism:
According to the Apex README and Macrocosmos docs, each Competition runs as a series of Rounds. Miners use the Apex CLI to submit Python-based algorithms. Validators continuously evaluate those submissions against benchmark tasks (game outcomes for Battleship, throughput and stability metrics for the Iota Simulator) and convert results into miner weights, which become emissions under Bittensor's Taoflow model.
RL Battleship asks miners to train and submit models that play Battleship on a 10x10 grid, scoring on how few turns they need to sink the opposing fleet. Iota Simulator goes further: miners submit routing and balancing algorithms that move forward and backward activations through a layered pipeline of heterogeneous nodes, with realistic constraints like node churn, variable latency, and bandwidth limits. The simulator mirrors the actual engineering challenges of Macrocosmos' separate IOTA training stack on SN9, which makes Apex a useful proving ground for ideas before they hit production training runs.
Development activity is unambiguous. The macrocosm-os/apex repository has 22 contributors and was last pushed on 2026-05-20, two days before this snapshot. The team is cutting weekly releases: v4.1.32, v4.1.31, and v4.1.30 all shipped between 2026-05-13 and 2026-05-15. The codebase is Python, MIT licensed, and the merge queue is being run by Macrocosmos staff (ewekazoo is the most recent merger). Apex is led by William Squires (CEO) and Steffen Cruz (CTO) at Macrocosmos.
The on-chain picture is more nuanced. Apex's emission share sits at 0.00% with seven-day net flows of -81 TAO, which under Taoflow means the subnet is currently outside the rotation receiving daily TAO injection. Price is 0.01006 TAO and market cap is roughly 50,093 TAO, with the pool holding 28,190 TAO. Root proportion is 0.16, so most of the pool is organic stake rather than protocol subsidy. Active miners count three, which is low even after accounting for the fact that competition-style subnets need fewer concurrent participants than inference subnets.
- Deregistration risk: With emission share at 0.00% and seven-day flows negative, Apex is currently not receiving emissions under Taoflow. If net flows stay negative, the subnet remains in the cold-pool state and faces continued pruning pressure on miners.
- Thin miner base: Three active miners is a small competitive field. Strong algorithms get rewarded, but the headline narrative of "open competition" needs more participants to play out.
- Competition for Macrocosmos attention: Macrocosmos operates five subnets including SN9 (Pretraining, IOTA) and others. Apex needs to keep proving its slot in that portfolio against subnets with clearer flagship narratives.
- Brand vs flows mismatch: SN1 carries legacy weight and recognition, but Taoflow does not care about brand. The subnet's economics will follow whichever competition catches stakers' attention next.
Another subnet, unpacked.
