Avalon-M1

Avalon-M1 is a new model series released on October 23, 2026 by US AI company General Reasoning (gr.inc) for running "agent organisations": instances of one shared model act as a CEO and hired specialists, coordinating through a shared message board and workspace to tackle long-horizon tasks.

Alongside the models, the company shared its training recipe, RLCR (Reinforcement Learning from Collective Rewards), and a technical report. The first model, Avalon-M1-Mini (120B total / 12B active parameters), is still training; what has been published are early results.

On October 25 an X user summed it up: "General Reasoning just released early results from Avalon-M1, a model trained to run full teams of AI agents that share one collective reward."

How an agent organisation works

• CEO + specialists: one instance acts as CEO and hires other instances as specialists
• Shared message board and workspace: agents talk on a board and work in a shared file space
• Collective reward: the whole team is rewarded as one

The RLCR recipe

• Every agent in the organisation receives the same collective advantage
• No per-agent credit assignment and no frozen sub-agents: everything is trained jointly
• For contrast, the report notes Kimi K2.5's PARL approach froze subagents and trained only the orchestrator

Early results (Avalon-M1-Mini, post-trained from the same base as Nemotron 3 Super)

Evaluation	Model alone	As an organisation
IMO-ProofBench-Advanced	29.8%	37.5%
FrontierScience	55.9%	63.5%

• On EdgeBench, organisations start more slowly but overtake single agents after about 1.7 hours of derived latency
• Training produced hierarchies consistent with organisational theory, new ways of dividing work and more peer review
• Safety finding: refusal of harmful actions stays near-universal, but reporting to humans declines ("organisational insularity")
• A collective inference-time scaling pattern: communication compute intensity grows with RL compute

How to learn more

• Release page: gr.inc/releases/avalon-m1 (video and roadmap)
• Technical report: gr.inc/avalonm1.pdf, "Avalon-M1: Reinforcement Learning from Collective Rewards"
• Status: Avalon-M1-Mini is still training; no public weights or API yet. The RLCR recipe and results are public.

Who should follow it

• Researchers in multi-agent systems, collective intelligence and multi-agent safety
• Teams interested in having several AI agents split up long tasks (proofs, science, complex engineering)

General Reasoning previously launched OpenReward and BackSearch (search over a frozen web archive as of any past date).