Monday summary of significant developments on Affine (SN120): revenue, measurable progress, mechanism changes, partnerships, and material token or chain shifts. Quiet weeks say so.
Affine (SN120), week to 2026-11-09
Verdict: Latest kings match Qwen3.5-9B on the team's AIME benchmark; Affine Video torn down.
- On the team's re-run benchmark at a 64K-token budget, the 8 Nov king scores 87.6% on AIME 2022–26 and 29.7% on ArXivMath, against 88.0% and 25.9% for Qwen3.5-9B and 60.0% / 4.4% for the base. Team harness, not reproduced; last week's 32K-budget figures are not comparable. benchmarks.json · commit
- Affine Video, launched as the "first model release" on 20 Oct, was torn down on 3 Nov; video.affine.io now serves an unrelated agent-hosting page. The subnet's only user-facing product lasted two weeks. commit · video.affine.io
- Evaluation went parallel from epoch 843 on 7 Nov and 44 miner updates were promoted in the week; training output caps rose from 8K to 32K tokens and the gate to 64K. Pay stays 10% each to the last ten promoted UIDs, 10 of 876 paid on chain. commit · llms.txt · chain read 9 Nov, block 9,468,896
Affine (SN120), week to 2026-11-02
Verdict: Gate fix on 31 Oct unlocked 24 promotions; miner pay cut to ten UIDs.
- After the held-out gate's noise floor was removed at epoch 354 on 31 Oct, 24 miner updates passed in 47 hours, against none in the 61 turns before. The team's fixed 4,976-problem MATH holdout rose from 86.8% at epoch 292 to 94.4% at epoch 544 on 2 Nov. commit · k1.json · performance.json
- Affine now publishes a public maths benchmark table re-running every king and 15 reference models on one node: pooled AIME 2022–26 42.5% at base to 64.5% at epoch 540, MATH-500 82.8% to 95.8%. Team-run, and still below Qwen3-8B at 70.4% and 97.0%. benchmarks.json · commit
- Since 28 Oct each epoch trains one miner's uploaded file, and only the last ten miners whose update beat the head are paid, 10% each; on 2 Nov the chain pays 10 of 683 UIDs, down from 143, with UID 0 taking 49%. Earning now requires moving the model, not submitting rollouts. llms.txt · scores.json · chain read 2 Nov, block 9,418,430
Affine (SN120), week to 2026-10-26
Verdict: Affine Video launched 20 Oct; held-out maths above base; miner pay rewritten three times.
- On 20 Oct Affine launched Affine Video, a free text-to-video studio on the open-weights MiniMax-H3 model, paired with a contest in which miners train a sparse-attention selector. First product anyone can use; 20 free credits a day and no price, so no revenue. const · tao.media · studio
- The 32-task held-out maths score ran 19–27 of 32 over 20–22 Oct (median 23) against 18 for the base, and the team reports 28 at epoch 136 on 26 Oct. First window clearly above base in this run; single epochs still swing four points. network-data.json · 1t1b
- Miner pay was rewritten three times: ten UIDs took all miner emission on 21–22 Oct (one up to 51%), 179–304 were paid from 22 Oct, and by 26 Oct weights follow a new flip formula with 12 UIDs banned. Who earns changes at the operator's discretion from week to week. incentive history · scores.json · FLIP_INCENTIVE.md · taoleeh
Affine (SN120), week to 2026-10-19
Verdict: Maths run restarted on a new 9B base model on 13 Oct.
- On 13 Oct Affine archived the Qwen2.5-Math run that had scored below its base and began training MiMo-V2.6-Distill-Qwen-9B from a fresh optimizer; 63 epochs had run by 19 Oct. Miners now submit whole eight-attempt groups with a thinking budget of 8,192 tokens. commit · llms.txt · mining.json
- The first three MiMo updates score 19, 21 and 22 of 32 held-out problems against 18 for the base (13 Oct). The team calls this not statistically established; a 128-task comparison is pending. eval doc · commit
- Since 13 Oct a single validator (Affine's UID 0) sets all weights, registration is by proof-of-work with no TAO burn, and 442 UIDs are registered against a new cap of 1,000. 229 miners were paid on 19 Oct, against 252 a week earlier. llms.txt · extrinsic · chain read, 19 Oct, block 9,317,436
Affine (SN120), week to 2026-10-12
Verdict: Trained maths model scores below the untouched base model.
- On 750 held-out problems, checkpoint 32 scored 66.1% against 73.1% for the untrained base (7 Oct). Training so far has made the model worse; the run restarted on 8 Oct from an earlier checkpoint. commit · restart doc
- Miners are paid again: 252 UIDs earn incentive on 12 Oct, up from 0 on 6 Oct. Hourly miner weights replaced the single-UID payout; no miner gets more than 0.7%. commit · chain read, 12 Oct
Affine (SN120), week to 2026-10-06
Verdict: Affine's maths model showed its first real gain.
- Affine's maths model improved for the first time: 77 → 86 of 128 held-out problems, then 84 on the next checkpoint. First sign the 4 Oct training pipeline works; miners are still unpaid. CP12 · CP13