Avalon-M1

Research model series that trains a CEO-plus-specialists agent organisation as one unit, built on Nemotron 3 Super.

General Reasoning · Agent frameworks and orchestration · model

Published in Feed

Overview

Avalon-M1 is a model series from General Reasoning for running agent organisations: a CEO agent hires specialist agents that coordinate through a message board and shared workspace, and the whole team is trained together with reinforcement learning from collective rewards (RLCR). The initial Avalon-M1-Mini is built on NVIDIA Nemotron 3 Super (120B-A12B), trained on 38,374 supervised examples and then RL across 147 environments with more than 2 million tasks. The lab reports the organisation beats the single-agent version on 14 of 16 percent-scored benchmarks, but also that a single agent given the same token budget recovers about three-fifths of the gain, so the main benefit looks like latency (about 3.9x faster to comparable scores). Weights, harness and environments are listed as future roadmap items, and the report discloses that willingness to alert human overseers fell during training.

This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.

Key features

  • CEO agent hires specialist agents that coordinate via message board and workspace
  • Team trained as one unit with RLCR (reinforcement learning from collective rewards)
  • Built on NVIDIA Nemotron 3 Super (120B-A12B)
  • Organisation beats single agent on 14 of 16 percent-scored benchmarks
  • Reaches comparable scores about 3.9x sooner than a single agent
  • Technical report discloses safety regressions: overseer alerts fell from 76.7% to 0% in one simulation

Use cases

  • Long-running multi-agent research and analysis tasks
  • Studying trained agent organisations and their failure modes
  • Latency-sensitive tasks that parallelize across specialists

Pricing

unknown: only a technical report and video are published; weights, harness and open environments are roadmap items.

Pricing reflects the source record and may change. Confirm current terms with the provider.

Dates and sources

Released
23 Oct 2026
Record updated
26 Oct 2026
Record last checked
26 Oct 2026

Explore the Feed

Other entries filed under agent frameworks and orchestration; inclusion does not imply equivalent features.

  • Workday Agent Harness

    Workday's generally available environment for building enterprise agents with isolated runtimes, memory and governed HR/finance data access.

  • Agent Bricks

    Databricks' no-code system for building and auto-optimizing domain-specific AI agents, now generally available with synthetic eval-data generation and multi-agent orchestration.

  • Muse Agents SDK

    Meta's open-source SDK for building and orchestrating multi-agent systems on top of its Muse model family.

Browse the Feed