TokenRhythm Calls a 4B Model a Step Toward Self-Improving AI

TokenRhythm released NeoHorse-1-4B on Sept. 9, billing the 4-billion-parameter model as "an initial prototype on the path toward recursive self-improvement," a term for a system that can find its own weaknesses, generate better training data and repeat the cycle without a person redesigning it each time. The release itself completes one pass through that cycle, not a repeating one.

The model card describes NeoHorse-1-4B as a fine-tune of Alibaba's Qwen3.5-4B, trained with what TokenRhythm calls a routing harness: a system that assigns tasks to a pool of models, records which one handled each task and how it went, and feeds that capability data into the next round of training. On a ten-benchmark suite, TokenRhythm reports a 64.87 macro average, 5.93 points above the 58.94 base Qwen3.5-4B posts on the same tests, with the largest gains on agentic benchmarks including a 9.79-point jump on WorkBuddy Bench and 10.5 points on VitaBench, and a smaller 0.77-point gain on the function-calling suite BFCL v4.

What 'recursive' means here

The technical report, posted to arXiv on Sept. 8 by a 37-author team, describes the harness as "closing an evaluation-selection-update loop in which what the system learns to do shapes what it learns from next," and calls the current work "a path toward harness-mediated RSI across successive iterations." Both the model card and the paper are explicit that this release represents one iteration: an updated model can be returned to the harness for another round, but the paper does not report having run that second round yet.

MindStudio, an independent outlet that reviewed the release the day after its arXiv posting, reached the same reading, writing that the model "doesn't claim to do this autonomously" and that the current release "only demonstrates one such loop." That a second, outside read landed on the same caveat TokenRhythm's own paper states is itself notable: neither source stretched the framing further than the other.

The model, apart from the framing

Set the self-improvement label aside and NeoHorse-1-4B is a fine-tune with real numbers behind it: Apache 2.0 licensed, natively supporting 262,144 tokens of context and extensible to 1,010,000, and trained with exact and near-duplicate removal, evaluation decontamination, structural validation and what TokenRhythm calls six-dimensional semantic evaluation of its training data. It arrives in the same weeks that larger Alibaba-derived releases such as Qwen3.8-27B have dominated open-weight download charts, though at a fraction of the size and aimed at agentic and coding benchmarks rather than general knowledge.

Benchmark	Qwen3.5-4B	NeoHorse-1-4B	Change
WorkBuddy Bench	24.62	34.41	+9.79
QwenClawBench	38.47	44.68	+6.21
PinchBench	71.19	77.33	+6.14

TokenRhythm has not published how many models sat in the routing pool during this training run, nor said when, or whether, it will publicly report results from a second loop through the harness. Until it does, small on-device releases like OpenBMB's MiniCPM5-2B are the comparison this model actually has to win on, not the recursive-improvement framing attached to it.