With the amount of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time
With the amount of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time
What It Does
SN66 Ninja is designed as a competitive AI coding-agent subnet. Its mission is to drive the development of top-tier open-source coding assistants by economic competition. Ninja aims to make the best open-source coding agent in the world, using Bittensor’s incentives to force miners to build agents that outperform each other on real coding tasks. In practice, this means a perpetual king-of-the-hill contest: miners register AI agents trained to solve coding problems, and they duel one-on-one on real GitHub issues. Only the reigning king agent earns TAO rewards each epoch, so miners must continuously innovate or be dethroned. According to the official dashboard, the competition is explicitly phrased as survival of the fittest for code – live duels on authentic code tasks where the king takes all emissions. This unique loop ensures that the network’s output is essentially the code of the current state-of-the-art agent.
In terms of problem-solving, Ninja addresses the issue of trivial benchmark gaming. Many AI code benchmarks can be gamed by over-optimizing on fixed tests, but Ninja uses genuine coding tasks. Miners contribute by providing an AI model (the coding agent) and its predicted code patches, which are committed on-chain via a referenced GitHub repository. Validators then run the same tasks, scoring each agent’s output by a similarity measure. For example, the score for a submission is matched_lines / max(agent_diff, ref_diff) against a reference solution. In each round, the agent with the higher similarity wins, and after many rounds the one with the most round-wins dethrones the current king. Thus the miners’ contributions are their agents (and generated code) and the validators’ role is to evaluate and compare those agents on real coding problems.
The end result of this process is a continuously evolving coding agent. At any point, the king agent – whose code resides on public GitHub – is effectively the network’s product: it represents the current state-of-the-art solution to coding tasks. Intended beneficiaries include the AI research community and software developers seeking powerful open-source coding tools: because the winning agent and its duel artifacts are fully public, everyone in the ecosystem can learn from them. Within Bittensor, the subnet is meant to attract miners who specialize in AI for code and validators interested in code quality. What makes Ninja unique in Bittensor is this pure, winner-take-all coding competition format with real tasks. Unlike other subnets that may use generated data or reward many participants, Ninja channels all emissions (100%) to the best performer. This no participation trophy approach means agents truly have to beat the best every time, constantly raising the bar.
What It Does
SN66 Ninja is designed as a competitive AI coding-agent subnet. Its mission is to drive the development of top-tier open-source coding assistants by economic competition. Ninja aims to make the best open-source coding agent in the world, using Bittensor’s incentives to force miners to build agents that outperform each other on real coding tasks. In practice, this means a perpetual king-of-the-hill contest: miners register AI agents trained to solve coding problems, and they duel one-on-one on real GitHub issues. Only the reigning king agent earns TAO rewards each epoch, so miners must continuously innovate or be dethroned. According to the official dashboard, the competition is explicitly phrased as survival of the fittest for code – live duels on authentic code tasks where the king takes all emissions. This unique loop ensures that the network’s output is essentially the code of the current state-of-the-art agent.
In terms of problem-solving, Ninja addresses the issue of trivial benchmark gaming. Many AI code benchmarks can be gamed by over-optimizing on fixed tests, but Ninja uses genuine coding tasks. Miners contribute by providing an AI model (the coding agent) and its predicted code patches, which are committed on-chain via a referenced GitHub repository. Validators then run the same tasks, scoring each agent’s output by a similarity measure. For example, the score for a submission is matched_lines / max(agent_diff, ref_diff) against a reference solution. In each round, the agent with the higher similarity wins, and after many rounds the one with the most round-wins dethrones the current king. Thus the miners’ contributions are their agents (and generated code) and the validators’ role is to evaluate and compare those agents on real coding problems.
The end result of this process is a continuously evolving coding agent. At any point, the king agent – whose code resides on public GitHub – is effectively the network’s product: it represents the current state-of-the-art solution to coding tasks. Intended beneficiaries include the AI research community and software developers seeking powerful open-source coding tools: because the winning agent and its duel artifacts are fully public, everyone in the ecosystem can learn from them. Within Bittensor, the subnet is meant to attract miners who specialize in AI for code and validators interested in code quality. What makes Ninja unique in Bittensor is this pure, winner-take-all coding competition format with real tasks. Unlike other subnets that may use generated data or reward many participants, Ninja channels all emissions (100%) to the best performer. This no participation trophy approach means agents truly have to beat the best every time, constantly raising the bar.
Product / Build
Ninja is live and operational on the Bittensor Finney network. As of late April 2026, the first epochs of SN66 have begun. On-chain stats show a still-nascent network: only 1 miner has registered (out of 256 possible), with a few dozen validators staking in it. The subnet metagraph uses 256 UIDs (neuron slots), one mechanism for the king-of-the-hill rule, and allocates 100% of block rewards to miners (meaning effectively the single king agent receives the entire emission). The alpha token price is currently around 0.013 TAO (approximately $4). Overall network TVL is only a few million TAO (approximately 3.7M staked) at present, reflecting its very early stage.
The technical architecture follows Bittensor’s standard Subtensor chain with an AI evaluation overlay. At launch, miners use the Bittensor CLI to join by burning a small amount of TAO. Once registered, each miner runs their coding agent locally or on servers. The project documentation suggests forking a provided TAU agent harness to build the model, meaning that each miner’s code appears as a GitHub repo with an agent/ directory of model code. The agents use standard machine-learning frameworks to solve tasks off-chain. Validators then fetch or compute predictions for each agent on the fixed coding tasks. The scoring is done by comparing each agent’s output to a gold solution with a line-match formula. These steps (task distribution, model execution, similarity scoring, and recording results and ranks on-chain) form the data flow of Ninja. The subnet’s blockchain layer handles epoch ticks and weight updates normally, so miners’ stake weights are updated according to performance.
No centralized GitHub repository for Ninja’s core code is publicly available. In practice each active agent lives in its own repo. Miners were told to study or fork the TAU harness, so presumably the code follows the same layout as other TAO reference agents. Because of this, we cannot detail files or recent commits. No official open repo or whitepaper has been released for Ninja specifically. Key infrastructure is inherited from Bittensor’s open-source base. In summary: the existing build consists of the Bittensor mainnet plus the specialized evaluation scripts. The only public dashboard is by Project Nobi, not an official code artifact.
Measurements and stats: the live state shows minimal engagement thus far. Active Miners: 1/256. Active Validators: on the order of a few dozen (one tracker lists 23). Total stake in SN66 is approximately 3.7 million TAO. As noted, emission is 100% to the winner. The alpha token market price is approximately 0.013 TAO (approximately $4) with market cap under $80K. These figures mean Ninja is tiny by crypto standards. Aside from Bittensor itself, there are no known external APIs integrated; tasks appear to come from public code and baseline LLMs are used for comparison. In development terms, the subnet is essentially under active use (duels ongoing) with no separate version 2 announcements. One implied line of work is the proposed SWE-bench cross-subnet benchmark to compare toolchains, suggesting future collaboration with other coding subnets. The end users of this system are the developers who run the code agents, and anyone who would use the resulting open-source model. With no paywall or API, the subnet produces a public AI tool: if the winning agent is published, any coder could use it. Thus the subnet operates more like a community competition than a commercial product.
Product / Build
Ninja is live and operational on the Bittensor Finney network. As of late April 2026, the first epochs of SN66 have begun. On-chain stats show a still-nascent network: only 1 miner has registered (out of 256 possible), with a few dozen validators staking in it. The subnet metagraph uses 256 UIDs (neuron slots), one mechanism for the king-of-the-hill rule, and allocates 100% of block rewards to miners (meaning effectively the single king agent receives the entire emission). The alpha token price is currently around 0.013 TAO (approximately $4). Overall network TVL is only a few million TAO (approximately 3.7M staked) at present, reflecting its very early stage.
The technical architecture follows Bittensor’s standard Subtensor chain with an AI evaluation overlay. At launch, miners use the Bittensor CLI to join by burning a small amount of TAO. Once registered, each miner runs their coding agent locally or on servers. The project documentation suggests forking a provided TAU agent harness to build the model, meaning that each miner’s code appears as a GitHub repo with an agent/ directory of model code. The agents use standard machine-learning frameworks to solve tasks off-chain. Validators then fetch or compute predictions for each agent on the fixed coding tasks. The scoring is done by comparing each agent’s output to a gold solution with a line-match formula. These steps (task distribution, model execution, similarity scoring, and recording results and ranks on-chain) form the data flow of Ninja. The subnet’s blockchain layer handles epoch ticks and weight updates normally, so miners’ stake weights are updated according to performance.
No centralized GitHub repository for Ninja’s core code is publicly available. In practice each active agent lives in its own repo. Miners were told to study or fork the TAU harness, so presumably the code follows the same layout as other TAO reference agents. Because of this, we cannot detail files or recent commits. No official open repo or whitepaper has been released for Ninja specifically. Key infrastructure is inherited from Bittensor’s open-source base. In summary: the existing build consists of the Bittensor mainnet plus the specialized evaluation scripts. The only public dashboard is by Project Nobi, not an official code artifact.
Measurements and stats: the live state shows minimal engagement thus far. Active Miners: 1/256. Active Validators: on the order of a few dozen (one tracker lists 23). Total stake in SN66 is approximately 3.7 million TAO. As noted, emission is 100% to the winner. The alpha token market price is approximately 0.013 TAO (approximately $4) with market cap under $80K. These figures mean Ninja is tiny by crypto standards. Aside from Bittensor itself, there are no known external APIs integrated; tasks appear to come from public code and baseline LLMs are used for comparison. In development terms, the subnet is essentially under active use (duels ongoing) with no separate version 2 announcements. One implied line of work is the proposed SWE-bench cross-subnet benchmark to compare toolchains, suggesting future collaboration with other coding subnets. The end users of this system are the developers who run the code agents, and anyone who would use the resulting open-source model. With no paywall or API, the subnet produces a public AI tool: if the winning agent is published, any coder could use it. Thus the subnet operates more like a community competition than a commercial product.
Team Info
The core individual associated with SN66 Ninja is known only by a pseudonym. The subnet’s creator identifies himself in community channels as Arbos (Discord handle @Arbos, Twitter @arbos_born). Arbos is an AI researcher/developer known for building competitive Bittensor subnets – for example he created the Distil subnet (SN97) with a similar winner-take-all model. No full legal name or corporate affiliation is publicly provided. Arbos appears to drive the project’s development and communication; his quotes emphasize Ninja’s goals. Along with Arbos, the only other figure listed is an email contact address (vilflea) for the subnet owner. It’s unclear if vilflea is an alias for Arbos or a collaborator, but no separate identity details are given. No additional developers, researchers, or employees are publicly credited for Ninja.
A related persona is Project Nobi (Twitter @projectnobi_tao), which has built the live SN66 dashboard. While not part of the official team, Project Nobi’s involvement is notable – they provide real-time duel analytics and branding for Ninja. Aside from this community contributor, no other individuals or firms are explicitly tied to Ninja’s creation. The subnet was launched via the Bittensor Launchpad (called BitStarter) in late 2025, implying coordination with the core Bittensor team but no specific partnership announcements were made. Thus there are no known external investors or partners specifically backing Ninja. It appears to be primarily a solo or small-team effort by Arbos, supported by community validators and stakers.
In summary, publicly identified participants are: Arbos (Lead developer/creator, Discord @Arbos, X @arbos_born, with background in AI agent development); vilflea (Contact alias, exact role unclear); and Project Nobi (Community member providing the SN66 live dashboard). No other team members, corporate sponsors, or investors have been publicly named. The initial token distribution was handled by the Bittensor team’s Launchpad, but beyond that no outside firms or venture backers are announced. The community engagement is primarily via Discord and Twitter, where Arbos and others share updates.
Team Info
The core individual associated with SN66 Ninja is known only by a pseudonym. The subnet’s creator identifies himself in community channels as Arbos (Discord handle @Arbos, Twitter @arbos_born). Arbos is an AI researcher/developer known for building competitive Bittensor subnets – for example he created the Distil subnet (SN97) with a similar winner-take-all model. No full legal name or corporate affiliation is publicly provided. Arbos appears to drive the project’s development and communication; his quotes emphasize Ninja’s goals. Along with Arbos, the only other figure listed is an email contact address (vilflea) for the subnet owner. It’s unclear if vilflea is an alias for Arbos or a collaborator, but no separate identity details are given. No additional developers, researchers, or employees are publicly credited for Ninja.
A related persona is Project Nobi (Twitter @projectnobi_tao), which has built the live SN66 dashboard. While not part of the official team, Project Nobi’s involvement is notable – they provide real-time duel analytics and branding for Ninja. Aside from this community contributor, no other individuals or firms are explicitly tied to Ninja’s creation. The subnet was launched via the Bittensor Launchpad (called BitStarter) in late 2025, implying coordination with the core Bittensor team but no specific partnership announcements were made. Thus there are no known external investors or partners specifically backing Ninja. It appears to be primarily a solo or small-team effort by Arbos, supported by community validators and stakers.
In summary, publicly identified participants are: Arbos (Lead developer/creator, Discord @Arbos, X @arbos_born, with background in AI agent development); vilflea (Contact alias, exact role unclear); and Project Nobi (Community member providing the SN66 live dashboard). No other team members, corporate sponsors, or investors have been publicly named. The initial token distribution was handled by the Bittensor team’s Launchpad, but beyond that no outside firms or venture backers are announced. The community engagement is primarily via Discord and Twitter, where Arbos and others share updates.
Roadmap
No formal roadmap document for Ninja has been released, but key milestones are evident from announcements and community records. The first event was in late November 2025, when SN66 was selected as the inaugural project on Bittensor’s LaunchPad. During this launch, the Ninja subnet token sale was offered at a discount (roughly 50% off) to raise funding; the sale quickly sold out, securing capital to be used in a TAO buyback. This marked the token generation and initial distribution phase of Ninja.
The next major milestone was the commencement of the competition. According to the live dashboard and social updates, SN66’s duels went live on April 28, 2026. On that date, Ninja began head-to-head coding agent battles (in fact, launching alongside SN62 Ridges). This live launch exists in the public record with the dashboard labeling it as a Live coding agent competition. Since then, rounds of duels have been running continuously as planned. No separate version updates have been announced; the roadmap is simply the ongoing tournament.
Beyond these launch events, few additional milestones have been explicitly listed. One development is the proposal of a standardized benchmark (SWE-bench) to compare coding agents across different subnets. This suggests an intent to cooperate with SN62 or other code-focused networks in future. The subnet’s long-term vision is to keep iterating the competition until it yields an agent that can write, debug, and refactor code better than anything else. In practice, that means the roadmap implicitly includes continually updating the tasks, increasing model complexity, and pushing performance each month/epoch. No timelines or new feature rollouts have been publicly announced beyond the existing competition framework. Any significant news so far has come informally on Discord or community forums.
In summary, the realized vision for Ninja is as follows: it has successfully launched on schedule (BitStarter Q4 2025, live battles Q2 2026) and is now in an ongoing era of competition. The future goal is the accumulation of a cutting-edge open-source coding model through these duels. No further official milestones have been specified; community updates and the duel results themselves serve as the evolving roadmap. Throughout 2026, any progress will be measured by new dethronements, higher scores, and eventual improvement of the coding agent.
Roadmap
No formal roadmap document for Ninja has been released, but key milestones are evident from announcements and community records. The first event was in late November 2025, when SN66 was selected as the inaugural project on Bittensor’s LaunchPad. During this launch, the Ninja subnet token sale was offered at a discount (roughly 50% off) to raise funding; the sale quickly sold out, securing capital to be used in a TAO buyback. This marked the token generation and initial distribution phase of Ninja.
The next major milestone was the commencement of the competition. According to the live dashboard and social updates, SN66’s duels went live on April 28, 2026. On that date, Ninja began head-to-head coding agent battles (in fact, launching alongside SN62 Ridges). This live launch exists in the public record with the dashboard labeling it as a Live coding agent competition. Since then, rounds of duels have been running continuously as planned. No separate version updates have been announced; the roadmap is simply the ongoing tournament.
Beyond these launch events, few additional milestones have been explicitly listed. One development is the proposal of a standardized benchmark (SWE-bench) to compare coding agents across different subnets. This suggests an intent to cooperate with SN62 or other code-focused networks in future. The subnet’s long-term vision is to keep iterating the competition until it yields an agent that can write, debug, and refactor code better than anything else. In practice, that means the roadmap implicitly includes continually updating the tasks, increasing model complexity, and pushing performance each month/epoch. No timelines or new feature rollouts have been publicly announced beyond the existing competition framework. Any significant news so far has come informally on Discord or community forums.
In summary, the realized vision for Ninja is as follows: it has successfully launched on schedule (BitStarter Q4 2025, live battles Q2 2026) and is now in an ongoing era of competition. The future goal is the accumulation of a cutting-edge open-source coding model through these duels. No further official milestones have been specified; community updates and the duel results themselves serve as the evolving roadmap. Throughout 2026, any progress will be measured by new dethronements, higher scores, and eventual improvement of the coding agent.