Loading comparison…
HEAD TO HEAD
TokenRhythm: neohorse-jev and Nandakishor M: vega, compared on what they decide, where they run, what they cost and what their publishers report.
| Property | tokenrhythm/neohorse-jev | nandakishorm/vega |
|---|---|---|
| Summary | TokenRhythm's prefill-only decision model for agent workflows. Built on its NeoHorse-1-4B (a Qwen3.5-4B derivative), it predicts Choice, Noul and Score answers over application-defined options without generating text, with optional single-image input. | Nandakishor M's typed-decision model. A frozen Qwen3.5 backbone is read, never trained, and cut off mid-stack; a small trained physics engine settles each choice, score or yes/no question into calibrated probabilities, a conformal set and an abstain flag. 0.8B and 4B. |
| Decides | choice, score, noul, classify, route | choice, score, noul, classify, route |
| Architecture | neohorse | vega |
| Fine-tuned from | tokenrhythm/neohorse-1-4b | qwen/qwen3.5-0.8b |
| License | apache-2.0 | apache-2.0 |
| Availability | Open weights | Open weights |
| Hosted by | — | — |
| Input price | — | — |
| Decision accuracy | 75.3% | 55.6% |
| Calibration error | — | 0.151 |
| Valid action rate | — | — |
| Median latency | — | — |
| p95 latency | — | — |
| Evaluation suite | JevBench public set (231 items), vLLM, maker's run | His 11 October 2026 benchmark, published with its results and its harness (a notebook in vegaml): macro accuracy over 43 public tasks for the 0.8B, 0.717 for the 4B; with JevBench (231 items) and DecisionBench (1,075 rows), 3,841 decisions in all, seed 20261011. On his own set of 2,050 choice decisions over six intent workflows (esci, clinc150, appliances, banking77, when2call, apibank), one choice question each, which he has not published with its question wording, he reports 0.763 (calibration error 0.026) for the 0.8B with its public_intents adapter, 0.389 without the adapter and 0.803 for the 4B with it. He publishes no accuracy for the typed_decisions adapters |
| Latest version | 2026.09 | 2026.10.12 |
| Variants | backbone, SHA256SUMS, assets, dist, package, tokenizer, vision | adapters, backbone, 4b |
| Size of latest version | 8.5 GB | 10.5 GB |
| Files | 38 | 31 |
| Downloads | 3 | 108 |
| Stars | 0 | 0 |
| Tags | system-one, qwen, multimodal, vllm, sglang, 4b | system-one, qwen, frozen-backbone, calibrated, conformal, abstention, adapters, multimodal, long-context, 0.8b |
| Updated | Oct 7, 2026 | Oct 29, 2026 |
Figures are from each model’s manifest; accuracy and latency are what the publishers report, on their own suites and hardware. Add a third model.
neohorse-jev is from TokenRhythm and vega from Nandakishor M. Both have open weights you can download and run. Both answer choice, score, noul, classify and route questions. vega is the smaller model, at 800M parameters to 4.0B.
They report on different suites — neohorse-jev 75.3% on JevBench public set (231 items), vLLM, maker's run, vega 55.6% on His 11 October 2026 benchmark, published with its results and its harness (a notebook in vegaml): macro accuracy over 43 public tasks for the 0.8B, 0.717 for the 4B; with JevBench (231 items) and DecisionBench (1,075 rows), 3,841 decisions in all, seed 20261011. On his own set of 2,050 choice decisions over six intent workflows (esci, clinc150, appliances, banking77, when2call, apibank), one choice question each, which he has not published with its question wording, he reports 0.763 (calibration error 0.026) for the 0.8B with its public_intents adapter, 0.389 without the adapter and 0.803 for the 4B with it. He publishes no accuracy for the typed_decisions adapters — so the numbers do not rank them. Test both on your own labelled examples.
neohorse-jev: Free (open weights). vega: Free (open weights). Open weights cost nothing per call beyond your own hardware.
Yes, both: systemone pull tokenrhythm/neohorse-jev and systemone pull nandakishorm/vega download the weights.