TokenRhythm's first open-weight models. They are 4B and 9B agent models post-trained from Qwen3.5, not pretrained in-house, and released under Apache 2.0 on September 7, 2026. The report calls them "an initial prototype on the path toward recursive self-improvement".

The idea is that a deployed routing harness already contains an RSI mechanism. TokenRhythm's router sends each turn of user traffic to a heterogeneous pool of models. For each turn it records the predicted capability demand, the service tier chosen, the tool interactions and the outcome. Those records become training data:

  • Data admission: the records keep interleaved reasoning, tool calls and harness context, and pass through deduplication, decontamination, structural validation, a six-dimension semantic evaluation and scene/goal/outcome labeling.
  • Training: the routing signals order SFT into a three-stage curriculum and drive routing-guided on-policy distillation.
  • The loop: capability-level evaluation sets the next training mixture. The updated models can return to the router's pool, closing an evaluate–select–update loop. The report runs it once and leaves repeated iterations as future work.

On the report's ten-benchmark protocol, post-training raises the macro average from 58.94 to 64.87 at 4B and from 65.60 to 69.04 at 9B. The protocol covers QwenClawBench, WorkBuddy, PinchBench, VitaBench, BFCL v4, τ²-Bench, HumanEval, LiveCodeBench v6, IFBench and IFEval. The 4B model beats Nanbeige4.2-3B (62.31) on that average, and the 9B beats Muse Glimmer 30B (67.86). Most of the gain is on agentic benchmarks; LiveCodeBench, IFBench and IFEval at 9B are flat. Both models have a native 262K context.

In September, NeoHorse-Jev-4B adapted the 4B into a prefill-only decision model. It returns probabilities for a choice, a yes/no check or a rating in a single forward pass, for routing and tool selection. It averages 77.70 over six decision benchmark groups. NeoHorse-1-4B has about 2,560 HuggingFace likes, and the repository has about 1,700 GitHub stars. The routing harness itself is described in TokenRhythm's OpenSquilla report.

Model Details

Context window 262,144
License Apache 2.0
Base model qwen3.5

Benchmark Scores

Benchmark Score Mode
Ten-benchmark agentic average 69.04 9B
Ten-benchmark agentic average 64.87 4B
tau2-Bench 90.82 9B
BFCL v4 67.43 9B

Variants

Name Parameters Notes
NeoHorse-1-4B — Post-trained from Qwen3.5-4B; GGUF and MLX builds
NeoHorse-1-9B — Post-trained from Qwen3.5-9B; GGUF and MLX builds
NeoHorse-Jev-4B — Prefill-only decision model built on NeoHorse-1-4B (Sep 23)

Paper

agentspost-trainingsmall-modelsrecursive-self-improvementopen-weight

Related