> ## Content Index
> Fetch the complete content index at: https://www.tao.media/llms.txt
> Use this file to discover other available public pages before exploring further.

# Astrid Appoints Zils to Run Bittensor SN127 and Raises 400 TAO in Alpha Sale
- URL: https://www.tao.media/astrid-appoints-zils-to-run-bittensor-sn127-and-raises-400-tao-in-alpha-sale/
- Published: 2026-11-06T21:27:43.000Z
- Updated: 2026-11-06T21:27:43.000Z
- Description: Astrid appoints Zils to operate Bittensor SN127, retains its owner key, and raises 400 TAO through a locked alpha sale as model mining goes live.
- Author: Tristan Hillerich
- Tags: News, Subnets, Astrid Intelligence PLC

Astrid Intelligence PLC (AQSE: ASTR) has appointed Zils, the decision-model business founded by Ken Nicholson, to operate its Bittensor Subnet 127 as an Astrid project. Astrid retains the subnet's owner key and will recognize its revenues and costs, separating day-to-day operation from ownership of the network.

The company also sold 200,000 SN127 alpha tokens to investors for 400 TAO in an over-the-counter transaction. The buyers agreed to lock the tokens, which will unlock on an agreed schedule. Astrid disclosed both developments in a [November 4 regulatory announcement](https://www.aquis.eu/stock-exchange/announcements/6004506?ref=tao.media).

Zils's live product on the subnet uses small models that choose between defined answers rather than generate free-form text. Its miners compete to improve shared open models, and Zils combines selected contributions into candidates for future product releases.

## Astrid Retains Ownership and the Subnet's Economics

The company's regulatory notice concerns the operation of SN127, rather than a change to Bittensor's protocol.

The token transaction was priced at 0.002 TAO per alpha, bringing 400 TAO into Astrid's treasury. Alpha is the subnet-specific token used within Bittensor's individual networks. Following the sale, Astrid held approximately 3,000 TAO-equivalent of SN127 tokens as of the announcement date.

"Subnet 127 now has a live product and a founder who builds in the open," Astrid CEO Siam Kidd said. "Selling part of our holding to investors who have agreed to lock it brings TAO into the treasury and puts committed holders behind the subnet at launch."

Zils began publishing validator weights on Bittensor mainnet SN127 on November 3, following earlier testnet work. Its [mainnet changelog](https://docs.zils.ai/changelog?ref=tao.media) records the launch of a competition organized around separate pools for the z3 model family.

## How Miners Build Zils's Shared Models

Zils's [decision-model workflow](https://docs.zils.ai/?ref=tao.media) starts with context, a question, and a set of possible answers. The model assigns probabilities to those answers, and the application uses the results to route a request, flag a case for review, or select an agent's next tool. The application retains control over permissions, actions, and human review.

The subnet's training competition improves the models underneath those decisions. Miners submit fine-tuned language and vision weights compatible with a fixed model configuration. They upload the weights to Hugging Face and record a commitment identifying the model on Bittensor; the weights themselves are not stored onchain. Zils publishes its [miner and validator software](https://github.com/ooo-hq/zils?ref=tao.media).

Validators test eligible submissions on private exam items and build what Zils calls a "soup," an equal-weight average of up to five model submissions. Starting with the best individual model, they add another only if the combination improves the score. Each member's contribution is then measured by removing it and checking how much performance falls.

Under the [launch rules](https://docs.zils.ai/operations/soup?ref=tao.media), miners earn seats for contributions that clear the required improvement threshold. A separate reward goes to an individual submission only if it beats both the existing house models and the combined soup by a specified margin. House models receive no payment, and unearned allocations are burned.

The four pools (z3-lite, z3-mini, z3-flash, and z3) each receive 25% of the base-emission budget. Lite and mini are running rounds; flash and z3 are scheduled to open later, with their allocations burned until then. The [live miner dashboard](https://zils.ai/miners?ref=tao.media) shows submissions, round results, and the models included in each pool.

A nightly combination is not automatically deployed to customers. At release windows every 14 to 28 days, a candidate must beat the current release on separate sealed tests, retain its advantage in a confirmation test 12 hours later, and pass checks on task-specific adapters. An adapter is a small set of learned changes that specializes a base model for a particular job. These checks connect the miner competition to versioned releases without changing customer models after every round.

## Hosted Decisions, Pricing, and the Evidence So Far

Zils is offering early access to hosted z3-mini decisions through an API, TypeScript SDK, and command-line tool. Customers can also train models using labeled text or photo examples, with evaluation before activation. Deployed customer models remain pinned to their base until an upgrade is requested. An experimental SDK feature ranks two to 10 options against a supplied criterion.

Decisions cost $0.06 per million input tokens, with output free. Training costs $0.20 per million training tokens, subject to a $2 minimum per run, using prepaid credit. Hosted decisions currently run on Zils's own platform GPUs; subnet miners improve the base models rather than serve customer inference. The company's [availability guide](https://docs.zils.ai/getting-started/roadmap?ref=tao.media) distinguishes that live workflow from planned miner serving and support for additional customer model sizes.

Zils's published research includes a customer-support experiment using the public ABCD dataset. A model trained on 1,024 conversations selected the recorded next action correctly in 396 of 500 held-out conversations, or 79.2%, compared with 353 of 500, or 70.6%, for hosted TypeSafe JEV 1.13.0.

That [comparison](https://zils.ai/model?ref=tao.media) used the earlier JevK5 4B base with a task-specific adapter, not the current z3-mini release. It measured next-action selection rather than resolution of a complete support conversation. Zils reused frozen predictions from its training study and tested JEV on the same inputs and choices. The result is one task-specific experiment, not a general performance claim for customer models; the research adapter has not been deployed.

Training also has a practical data-handling requirement. Examples are copied to approved workers, whose operators can read and retain them. Zils explicitly says the service is not confidential computing, so customers need authorization to share the material they upload.

## A Scientific Claim-Review Pilot

Zils has also [announced a pilot with DeSciClaims](https://x.com/i/status/2117283819443171785?ref=tao.media), applying a custom decision model to scientific claim review. High-confidence decisions pass automatically through the partner's data pipeline, while uncertain cases continue through its existing validation process.

![](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/11/Zils-and-DeSciClaims-pilot-graphic-_-Zils-on-X.jpeg)

Zils and DeSciClaims pilot graphic | [Zils on X](https://x.com/i/status/2117283819443171785?ref=tao.media)