{
  "slug": "2026-10-05-bankr-founder-trains-on-device-model-on-anonymized-agent-sessions",
  "title": "Bankr founder says on-device model is being trained on anonymized agent sessions",
  "date": "2026-10-05T17:40:54.000Z",
  "updated": "2026-10-05T17:40:56.146Z",
  "tags": [
    "agents",
    "infrastructure",
    "tokens"
  ],
  "summary": "The small model aims to keep user conversations with BankrBot agents on-device, with no cloud LLM access to trades.",
  "kind": "article",
  "body": "# Bankr founder says on-device model is being trained on anonymized agent sessions\n\n*The small model aims to keep user conversations with BankrBot agents on-device, with no cloud LLM access to trades.*\n\nBankr founder @0xDeployer said on October 5, 2026, that he is training a small model to run on a user’s device using tens of thousands of anonymized BankrBot agent sessions.\n\nIn a post on X at 00:03 UTC, he said the goal is to keep a user’s conversation with their agent on their phone, with no cloud LLM reading their trades. He did not disclose the model’s name, parameter count, or compatible devices, but said he would share performance results.\n\nBankrBot is a trading agent framework on Bankr, a platform that distributes 95% of a 0.7% pool swap fee, 0.665% of trading volume, to token creators, according to the company’s documentation. Creators can claim fees at any time, and the platform vests 15% of supply to them over one year with a 30-day cliff. It also enforces a 2% wallet cap for the first five minutes after launch.\n\nThe training data comes from agent sessions, which record conversations and actions when a BankrBot agent runs. The post states the sessions are anonymized but does not describe the method, retention period, or where training occurs.\n\nThe announcement follows two other Bankr experiments this month. On October 2, 2026, @0xDeployer began a 10-day trading challenge in which a BankrBot agent managed trades without human intervention. The agent posted its first take-profit hit the same day, following a dry run with a 75% win rate, according to his post at 21:01 UTC. The live trade opened at base address 0xfb31f85a8367210b2e4ed2360d2da9dc2d2ccc95. The post did not disclose the initial trade size, the asset traded, or performance beyond the dry-run win rate.\n\nOn October 1, 2026, he said almost 25% of the BNKR token supply was staked, up from 15.4% on September 27, 2026. Bankr opened early staking on September 26, 2026, the same day he wrote that the company had bought back seven figures of BNKR to stream to stakers. According to Bankr’s staking documentation, rewards are paid in BNKR and streamed weekly, with each week’s batch released over seven days. A staker’s share depends on total weight staked.\n\nThe October 5 post did not say when the model would be released, whether it would be available to users outside Bankr, or how it would be evaluated.",
  "sources": [],
  "wordCount": 404,
  "readingTime": 3,
  "canonical": "https://protocolflow.today/a/2026-10-05-bankr-founder-trains-on-device-model-on-anonymized-agent-sessions",
  "alternates": {
    "md": "https://protocolflow.today/a/2026-10-05-bankr-founder-trains-on-device-model-on-anonymized-agent-sessions.md",
    "json": "https://protocolflow.today/a/2026-10-05-bankr-founder-trains-on-device-model-on-anonymized-agent-sessions.json"
  },
  "translations": [
    {
      "lang": "ja",
      "url": "https://agentkeizai.today/a/2026-10-05-bankr-founder-trains-on-device-model-on-anonymized-agent-sessions",
      "name": "エージェント経済"
    },
    {
      "lang": "ko",
      "url": "https://agentpost.today/a/2026-10-05-bankr-founder-trains-on-device-model-on-anonymized-agent-sessions",
      "name": "에이전트포스트"
    },
    {
      "lang": "zh-CN",
      "url": "https://lianzhi.today/a/2026-10-05-bankr-founder-trains-on-device-model-on-anonymized-agent-sessions",
      "name": "链智日报"
    }
  ]
}