Bittensor Weekly — October 2, 2026

The pitch, from Const on day one in Montreal: a subnet burns some of its alpha into gamma, a transferable usage credit, then spends the gamma on compute, inference and other services from sibling subnets. Minted at a stable oracle price, so $5 of alpha buys 5 gamma and the count never changes after that. Const and Keith Singery talked it through in a 36-minute fireside on day two.

The mechanics are strict. Gamma is a one-way trip: it never converts back to alpha, and only the subnet owner key can accept it. Value capture flows through the subnet's own token on every leg. Owners can redirect part of miner emissions to buy other subnets' gamma automatically, which lets a product team pull inference, compute and storage from across Bittensor without raising outside capital. The credit is built to travel beyond the network's walls if outside providers accept it.

The pattern underneath already exists: Lium, Chutes and Targon run buyback-and-burn credit models on private databases. Gamma puts that model at the base layer, one atomic transaction from any chain, no email and no API key required.

The market read it fast: TAO gained 3.75% on the proposal. The root channel has spent two days on second-order effects since, from composability arguments to the price impact of rerouting emissions into gamma. The room's best detour: app subnets with no miners of their own, paying their whole budgets upstream in gamma. That is the customer shape the proposal is reaching for.

KuCoin's write-up called it the biggest change to Bittensor's token design since dTAO. Nothing is voted on. The proposal has to prove itself to the room first, same as the Burnweight proposal did.

Hone (SN5) shipped V4, then turned the difficulty up.

V4 landed Tuesday at 16:00 UTC: every task is an open-source repo with bugs planted inside, plus a short bug report. A miner works the repo, returns a Git diff, and hidden checks grade it offline. Failed runs come back with a signed notice naming the exact test that broke. This is the update meant to get Hone's first model out.

Then came Thursday's fix, after complaints that new miners were getting pruned before they could build a record: the problem rate went up 4x and the immunity window (the head start before deregistration) stretched from about 17 hours to 24. Each miner should see roughly 50 problems before it runs out.

Harder problems, longer runway. The first model is what all of it is for.

Targon (SN4) and IOTA (SN9) widened the compute stack at the summit.

Targon launched Bare Metal and Sandboxes: an entire physical machine with no hypervisor (no layer between the operator and the hardware) for the teams that want it, and disposable Linux environments that boot in seconds, fork for parallel agent runs, and stop billing while paused. That pushes the subnet past confidential containers, and both products are live now.

IOTA brought the other half: the iota SDK, a few lines changed from standard PyTorch, over Liquid Compute, which pools scattered GPUs, reissues work when machines drop, and refuses to wait forever on the slowest box. Summit numbers put it at up to 10 trillion parameters and up to 3x lower cost than centralized labs, with commercial launch in the coming weeks.

Machine rentals, agent sandboxes, scattered-GPU training: the compute stack under everything got wider this week.

Hot. Thursday ran 1,019 messages across 70 channels, against 268 the day before. Score (SN44) led at 96, after two new detection tasks went live. ORO (SN15) 93. Ditto (SN118) 61. Beam (SN105) 42.

Building. Claims (SN111) took V1 live with its consensus layer: panels of miners review the hardest cases, and their agreement mints each Gold record into the database. Cascade (SN91) opened Ephemeris to public traffic: six forecasting models behind one request format, pay only for the models that ran; one miner fed it Spotify history and got a 2027 Wrapped forecast out. Cortex (SN100) shipped SDKs and a Claude plugin for OpenType, its open answer to Jev: 250-400ms responses on a single GPU, with 70% of emissions riding on miners making the model faster. MVTRX (SN79) took proportional pay live on mainnet: one strategy spread across many accounts now earns what running it once earns. Apex (SN1) landed its seventh Humanoid Olympics event, the 200m race walk, with foot contact checked at every physics step. Hippius (SN75) shipped Shared Drives, encrypted end to end, with the storage owner footing the bill. TPN (SN65) shipped Fluid Bench: versioned knowledge benchmarks that run private and freeze before scoring, so submissions can't be tuned to the exact test. Ridges (SN62) opened Test Generation, splitting emissions 48% tests, 48% database engineering, 4% linting.

Provenonce (SN87) ran its first full cycle on the public testnet: a validator graded two miners across 16 sealed cases and 32 signed responses, set weights at block 8,102,381, and read the result back from the chain. Only the weights went on chain, and the team flags it as one synthetic loop with no independent miners yet. engy (SN53) cut Cloudflare out of its miner connection path, and the before-and-after is lopsided. 404-GEN (SN17) previewed how closed models will work in 3JS: GPT for image-to-3D as part of a miner's submission, with validators replaying a recorded transcript instead of re-calling the API, and keys proven with a zero-knowledge proof (proof the credential works without revealing it). Pareton (SN10) and engy put their partnership live: a Pareton-tuned Qwen3.8 is serving production traffic on engy's rails. Verathos (SN96) and Nodexo (SN106) merged into one network: pooled GPUs with proofs on top, and external revenue that buys SN96 alpha and locks it instead of selling. BitMind (SN34) swapped fresh holdouts into its deepfake tracks and re-scored the leaders against them, raising the bar for anyone chasing.

Cold. The map stays lopsided: a dozen teams carried the week's news, and the long tail beyond them stayed where it always is. Same shape as most weeks.

Friction. Affine (SN120) ran the week's sharpest set of changes: the dethrone margin cut from 0.20 to 0.10 sd (standard deviations: how convincingly a challenger must beat the king) on a directive straight from Bittensor co-founder Jacob Steeves, the scoring model swapped to GLM-5.3-Flash with every score rebaselined, and then distillation dropped entirely, with the team apologizing for not addressing the miner impact when the change was first announced. Miners who trained models for the old competition got a two-day window to submit under the old scoring system.

Reliquary (SN81) is between tasks and its burn is eating the week. "There is nothing to earn right now," one miner wrote on the way out, "mining is currently unprofitable and GPU costs keep adding up." The team's answer is more work: four fine-tuning jobs went live Wednesday covering 20% of emissions, and RomAI says more task types are coming, each one lowering the burn. Miners keep asking when.

Instant (SN46) spent the week fixing bugs a community member found, and said so every time. Kevin filed reports detailed enough that both issues he flagged checked out: a partial UTF-8 character at the end of a stream could restart an entire GPU backend, and one bad proof was zeroing busy miners. Both fixed, and the fixes shipped Thursday in Miner v0.6.0 with a fairer proof-tolerance rule attached.

Then the thank-you. The team's writeup credited Kevin by name, and Seby added the line that made the channel: "Shout out to @Kevin for the good faith support. This is what Bittensor is all about."

No campaign around it. A community member filed detailed reports, the subnet fixed its bugs and said thanks in public. Some weeks the whole loop is that short.

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]