TokenRhythm Open-Sources NeoHorse-1: 4B/9B Models Trained on Agent Trajectories

TokenRhythm (the OpenSquilla team) released and open-sourced NeoHorse-1, its first validation of the "learning from agent runs" idea: 4B and 9B checkpoints post-trained from Qwen3.5, with a native 262K context window and a technical report on arXiv. The motivation is a familiar pain point: agents can find and fix bugs through trial and error, yet repeat the same mistakes the next day—conversations have memory, but the underlying model never improves, and valuable trial-and-error work is discarded when a task ends. TokenRhythm aims to systematize this run experience into training data that flows back into model weights. Early official results suggest the direction works: NeoHorse-1 4B beat its Qwen3.5 base on all ten tests, raising the average score from 58.94 to 64.87.

Confirmed

• TokenRhythm has three parts: OpenSquilla runs tasks, the TokenRhythm API connects tasks to different models, and NeoHorse-1 learns from the experience.
• Step one is Data-RSI: the system observes the full execution process—which model did the work, which tools were used, where it failed, how it recovered, whether the final answer was valid—and useful run logs become training samples for the next model generation.
• Step two is a test-driven data loop: models are tested to find weaknesses (e.g., struggling with coding, tool use, or instruction following); the next training batch increases the share of such tasks, and the updated model is tested again, repeatedly.
• NeoHorse-1 is open-sourced and the team calls it an early prototype toward recursive self-improvement (RSI). rohanpaulai's coverage confirms the framing: most agent systems discard the most valuable training data after each task—routing records, tool calls, failures and recoveries—which TokenRhythm recycles.
• Tomar cites official early data: NeoHorse-1 4B outperformed its Qwen3.5 base on every one of ten tests, average score up from 58.94 to 64.87.

Unconfirmed

• The ten-test results come from TokenRhythm's own early data, as Tomar himself notes; no independent third-party verification yet.

Why it matters

• Agent run experience today mostly stays at the conversation level and never reaches model weights. TokenRhythm offers a path to systematize real agent execution traces into training data, and the open-source NeoHorse-1 with across-the-board gains over its base provides the first reproducible public validation of the "learning from agent runs" thesis.

2026-09-16 ~ 2026-09-17 · 11 related posts

• AI agents fix bugs then forget — TokenRhythm turns agent runs into model training data — PrajwalTomar_ · 2026-09-16
• TokenRhythm's bet: train models on lessons from your agent's own runs — PrajwalTomar_ · 2026-09-16
• Inside TokenRhythm's three-part stack: OpenSquilla, a model API, and NeoHorse-1 — PrajwalTomar_ · 2026-09-16
• Data-RSI: turning useful agent runs into training examples for the next model — PrajwalTomar_ · 2026-09-16
• TokenRhythm's loop: test for weak spots, weight next training batch toward them — PrajwalTomar_ · 2026-09-16
• [source] NeoHorse-1 debuts: Apache 2.0 4B/9B models trained from agent runs, 262K context — PrajwalTomar_ · 2026-09-16
• [source] NeoHorse-1 4B beats its Qwen3.5 base on all ten tests, average score up 58.94 to 64.87 — PrajwalTomar_ · 2026-09-16
• [source] NeoHorse-1 open-sourced: 4B/9B Qwen3.5 post-trained models targeting recursive self-improvement — PrajwalTomar_ · 2026-09-16
• TokenRhythm Launches NeoHorse-1: 4B/9B Models Post-Trained on Agent Execution Traces — rohanpaul_ai · 2026-09-17
• TokenRhythm releases NeoHorse-1: 4B/9B models post-trained on agent execution traces — rohanpaul_ai · 2026-09-17

1 near-duplicate retellings: PrajwalTomar_