Bittensor Weekly — November 6, 2026

Iota (SN9) shipped Decide 4B, the first public model out of Macrocosmos: a 4B decision model under Apache-2.0 that answers with probabilities instead of text.

A decision model takes a state (a support ticket, an agent trace, a policy plus a request, any JSON) and a set of typed questions: is this true, which option applies, which level. It returns a calibrated probability for each answer. No prose, no prompt engineering, about 33 ms a request on one GPU, roughly 9 GB of weights in bf16, and it runs on CPU too.

The number that got passed around all week: 78.6 on JevBench (a benchmark for decision models), against 62.4 for OpenAI's Decisions API.

Then the footnote on the model card, which most people quoting that number will skip. 78.6 is Macrocosmos' own run on the 231-question public split. The other models' scores are public leaderboard numbers. The evaluation on the sealed set is pending, and the card says so.

The training run is the part worth staring at. Sixteen RTX A4000s, one card per machine, merged with streaming DiLoCo (training that syncs fragments on a schedule instead of waiting for every node at once). LoRA adapters on Qwen3.5-4B, rank 64, 750 steps, roughly 879,000 deduplicated records, no cluster anywhere in it.

Fine-tuning a 4B model for routing tickets will not win anyone a Nobel. It turns the "train on borrowed GPUs" pitch into a downloadable artifact with a benchmark score and a working example, which is further than that pitch usually gets.

The weights are free and the request format is a drop-in for existing decision-API calls. Run it on your own traffic before you believe any score in this issue, mine included.

Ditto (SN118) deleted the clause that destroyed half of every miner's reward.

The split is now 18% owner, 33% miners, 49% validators and stakers, live on-chain since the start of the week, up from 41% for the validator column. The miner slice looks smaller on paper, 41 down to 33. Half of that old slice used to go to burn, so the share that reaches mining goes from about 20% to 33%.

Two things follow. More TAO stays in the pool, which deepens the market for anyone buying or selling SN118 alpha. And the subnet stops paying miners in a token it was destroying at the same time, an arrangement that was always hard to explain to anyone outside Bittensor.

A burn promises the supply gets smaller. A payout promises someone gets paid. Ditto picked the second one, and the validator column picked up eight points in the rebalance.

Apex (SN1) rebuilt its robot assembly round and wiped the leaderboard doing it.

v0.5.1 landed on November 1. Same task, same scoring, same interface. The peg and hole estimates (where the part sits versus where the slot is) now carry 5 to 11.5 mm of offset, against 1.5 to 5 mm before. The best hand-tuned scripts from v0.5.0 scored 0.85 to 0.94 under the old build. Under the new observations the same entries land between 0.09 and 0.27.

That is the whole trick of an optimization subnet, executed in public. A script that wins is a script that memorized the noise. Widen the noise and the script has nothing left. Rounds 15 and 16 stay on the old build, and the incumbent gets re-scored in round 17.

Apex also opened its URnetwork competition. The entry is one file applied to a live product's connect path, built offline, run nine times against a control frozen for the epoch, and the winner merges into URnetwork's codebase as the next epoch's baseline. Six epochs, seven days each, 20% of Apex emissions on it.

Hot: the queues and the arguments.

Affine (SN120) ran the loudest channel of the week at 130 messages, and the complaint was queue fairness. 250 registrations were waiting, only about 15 had a valid batch ready, so every turn skipped past roughly 100 idle keys while miners who paid 1 TAO per UID and rented GPUs waited. The fix recomputes expected wait on every turn: one day minimum, or the number of registrations ahead of you divided by turns per day. The competition underneath now holds 19,305 problems, about 500 drawn per round, with the pass bar decaying from 1.0 to 0.25 over 12 hours so the gate cannot sit closed.

Niome (SN55) came second at 128 messages, and the subject was the seed. One miner reproduced the published top-10 scores to four decimals locally and asked where the seed comes from. The team's answer: validators generate it in the validation stage, it reads zero during broadcast, and a submission that misses that window is ignored. Nobody in that channel sounded convinced, and the self-mining questions kept coming.

Cacheon (SN14) at 97 messages was the best read of the week. The owner is deciding what hardware the next kernel arena runs on. B300s are expensive, and the worry is miners tuning kernels nobody will buy. H200 is the likely next step, 8x RTX PRO is on the table, and the condition for either is miners producing something a customer would pay for.

Score (SN44) at 79 messages was a wall of ground-truth complaints, block by block, with miners posting screenshots to prove a label was wrong. Six challenge blocks were pulled in one sweep earlier in the week for exactly that, and the tiebreak similarity threshold moved from 0.045 to 0.041 on Wednesday.

Building, shipping, patching.

ComputeHorde (SN12) took a security review and published it. Dovetail researchers broke job authentication, the isolation between one job and the next, and the handling of job artifacts, with reproducible attacks attached. Everything they found is fixed and the patch is public. A path traversal bug (a way to read files outside a job's own folder) reported through GitHub in July had sat unnoticed until this review. A subnet that runs other people's code for a living cannot treat that report queue as optional.

Instant (SN46) onboarded its first paying Stripe users with no funds settled yet, opened crypto payments to API users, and pays its first buyback and burn Friday at 17:00 ET on two public coldkeys.

Blockmachine (SN19) split miner emission across the 11 chains it serves. Bittensor gets 15%, BSC and Polygon 17.5% each, Ethereum and Base 10%, Robinhood 7.5%, Mantle, Optimism, Arbitrum and Arc 5%, Scroll 2.5%. Cost and demand set the order, which is a defensible way to run it. The announcement gave miners about two days, and the miners said so.

Gradients (SN56) flagged two Round 2 miners as duplicates, then reversed it. The dedup check through the Claude API saw identical hyperparameters down to the decimal. One miner had left those numbers in an uploaded artifact, the other admitted reusing them from last week. The code and training logic differ, both attempts stand, and every future dedup call now needs two reviewers, public hotkeys and a 24 hour window to argue.

Kokoah (SN39) took the slot with agents trading Hyperliquid inside sealed machines and a builder fee of 0.025% per order that buys and burns SN39 alpha. Mining starts in a few days. Trishool (SN23) settled challenge 15, pays the winning hotkey for a week, then goes to full burn until Phase 4 opens.

Cold: 25 channels stopped posting before October 30. Cascade (SN91) spent its messages arguing about rejected challengers. NeverPlayAlone (SN98) and Minos (SN107) spent theirs fielding self-mining questions, with one Minos miner asking how a subnet nobody outside can audit ends up near the top of the rankings.

Nova (SN68) had a bad week on infrastructure and told everyone the truth about it.

Two validator runs crashed. The dashboard fell a couple of epochs behind. Miners asked, twice, whether anything was running at all. The owner's answers were short and specific: infrastructure instability, fixes in progress, one run pinned and active, and no explanation for why the interface showed crashed. Then the logs got shared, which is more than most subnets do with a broken week.

A miner called luckyra, who had spent the week chasing it, wrote this:

Two crashed runs and a compliment like that in the same week is not a contradiction. It is what happens when a team answers the question instead of the tone.

Next week's issue goes out Friday. Subscribe and it lands in your inbox: https://newsletter.bittensor.quest

Shaping the next issue: tell me what to keep, what to cut, and what you want more of — mailto:[email protected]