Affine (SN120) – The RL Arena For Improved AI Reasoning

Most AI reasoning progress still happens inside closed labs. Affine is trying to change that.

Affine is Bittensor Subnet 120, an open reinforcement learning arena where miners compete to improve models on coding, tool use, and math. Only models that outperform the current champion across multiple environments earn the bulk of the rewards.

The project is led by Jacob Steeves (known as Const), co-founder of Bittensor himself. 

Winning models are made publicly available, turning competition results into usable open capability almost immediately.

In short, Affine is a live experiment in whether open incentives can drive real gains in AI reasoning faster than closed teams alone.

1. Quick Overview  

Affine Alpha (SN120) – Live Snapshot (as of ~19 Sep 2026 at 10:00 AM UTC)

• Purpose: Incentivized reinforcement learning competition for open reasoning models (coding, tool use, math, program abduction). Winner-takes-most style rewards for models that beat the current champion.
• Launch Date: 10 June 2025
• TAO price: ~$265
• Subnet’s Token Price (Alpha): ~0.0489 τ (≈ $13.16)
• Market Cap / FDV: Market Cap ≈ $47–55M range. FDV: $271.27M
• Emissions %: Varies by snapshot; recent figures shows ~6.59% of network emissions
• Root proportion: ~18.71%
• Volume/ MC: 0.36%

Note: Subnet metrics (especially emissions %, volume, and exact alpha price in τ) move every tempo. Always check live explorers for the absolute latest numbers.

2. TL;DR

• What is Affine
  Affine is Bittensor Subnet 120. It is an incentivised reinforcement learning environment focused on reasoning tasks such as coding, tool use, and math.

• How it works
  Miners improve open-weight models and submit them. Validators run duels against the reigning champion across multiple environments. Only models that dominate earn the bulk of emissions. Winning models become publicly available. Submissions are now private.

• Why it matters
  Closed labs improve reasoning on their own schedule and keep the best versions private. Affine turns continuous improvement into an open market anyone can enter. Results flow straight back into usable public models.

3. Product and Features

• Miners
  Train or fine-tune models, commit them on-chain, and compete to dethrone the current champion. Private submissions protect weights from immediate copying.
• Validators
  Generate challenges, evaluate models with the current scoring system (teacher-anchored Reason and Grounding scores), publish results and set weights. Scoring has been updated multiple times to close exploits.
• Core functionality
  Winner takes most (or short-window equal share) emissions. Continuous duel queue. Public crowned models. Benchmark tracking. Integration with other subnets for hosting. Live reign and duel data shown on the homepage.

4. Moats

• Strong founder alignment (Bittensor co-founder Const is directly involved).
• Explicit anti-gaming design (Sybil-, copy-, decoy- and overfitting-resistant claims).
• Private submissions plus public crowned models plus delayed checkpoint release for independent verification.
• Teacher-anchored scoring that tries to stay closer to real capability rather than pure LLM-judge gaming.
• Direct pipeline of improved models into other subnets (especially inference/hosting layers).
• High historical emission ranking and deep liquidity relative to many peers.
• Active research documentation and public mechanism evolution.

5. The Team Behind it?

• Founder and lead: Jacob Steeves (Const), co-founder of Bittensor.
• Operated under the Affine Foundation.
• Public contributions from a wider group, including strong Chinese AI developer involvement (stated in earlier write-ups).
• Contact: hello@affine.io

6. Code Quality

• Repo: https://github.com/AffineFoundation/affine
• Last Commit: 10 September 2026 
• Languages: Almost entirely Python. Uses a modern uv workspace setup with some Shell scripts for ops and installation.
• Stars / Contributors: Only 2 stars and 4 forks.
• Hardware: GPU 
• Forkable?: Yes. The repository is fully public and can be forked like any standard open-source project.
• Quality: Strong research and mechanism, detailed documentation, clear versioning. It is generally a perfectly functional code.

7. Social Sentiment

• Generally positive among technical Bittensor participants for the mechanism design, Const’s direct involvement, and the focus on reasoning over pure scale. 
• Community posts and guides highlight it as one of the more intellectually serious subnets. 
• Broader Bittensor discourse sometimes mixes praise for innovation with criticism of emission dynamics and extractive behaviour across the ecosystem; Affine itself is more often cited as an example of ambitious mechanism work. 
• Official communications emphasise open collaboration and “commoditising reasoning.”
• Official updates come from @affine_io.

8. Whale Activity

• Deep liquidity pool (tens of millions USD equivalent).
• Specific large-wallet movements are not exhaustively tracked in public summaries at the time of writing.
• 256/256 slots filled.
• Owner lock and large holder activity monitored on explorers.

9. Valuation Model

Is it undervalued?

You’re paying for real technical ambition and Const’s involvement rather than proven external usage. If the models keep improving and start seeing actual demand, the current level looks reasonable. 

Not screaming cheap, but no longer expensive either. 

If usage stays mostly internal, it’s fairly priced at best.

Whales and support

Liquidity is deep and sticky, which usually means larger players are comfortable holding. 

Holder concentration is visible and the pool doesn’t look fragile.

Overall, Affine has stronger structural backing than most subnets at this market cap. The question is whether the reasoning models become useful outside the arena itself.

10. Technical Chart Snapshot

• Trend: Clear multi-month downtrend from the $29–$30 ATH zone
• Current Phase: Low-volatility consolidation after the selloff, sitting near recent lows
• Support: $10.50–$10.80 (recent swing lows)
• Resistance: $12.50–$13.50 (prior bounce area and volume shelf)
• Volume: Moderate to weak on recovery attempts (buyers not aggressive)
• Note: Needs a strong reclaim of $12.50+ with rising volume to suggest any real bounce; loss of $10.50 risks further downside.

11. Thesis

Bullish

• Robust scoring plus continuous public model gains could make Affine a core open reasoning engine. 
• Founder involvement and emission history support the case.

Bearish

• Incentive mechanisms can be gamed. Commercial traction is still hard to measure. 
• Price trend weakening; Abrupt price dump post-launch 
• Slow recovery.

12. Final Verdict

🟡 Speculate

• Constructive but selective. Affine is one of the higher-conviction mechanism experiments on Bittensor precisely because of its founder and technical ambition. 
• Suitable for those who believe open, incentivised reasoning markets can compound; size accordingly given the still-early product-adoption stage and typical subnet volatility. 

This is not financial advice; do your own research and size positions carefully.

13. Glossary

• SN120 — Affine’s subnet number on Bittensor.
• Alpha / SN120 token — The subnet’s native token.
• TAO — Bittensor’s native token.
• Miners — People who improve and submit models.
• Validators — Nodes that evaluate models and set weights.
• Dominance — Outperforming across the full set of test environments.
• Chutes (SN64) — Hosting subnet used for public model inference.
• RL — Reinforcement Learning.

14. Sources

• Official site: https://affine.io 
• GitHub: https://github.com/AffineFoundation/affine
• Explorers: Taostats, GeckoTerminal SN120 pool, dTAOscan
• X: https://x.com/affine_io
• General Bittensor reference: https://docs.bittensor.com 

15. Disclaimer

All figures and rankings move quickly; cross-check live sources before any decision. This article synthesises publicly available information as of mid-September 2026 and is intended for educational purposes only.