System OneSystem One
All articles

September 19, 2026 · 3 min read

The open reproductions — Nimble, Decider, Von, JevK5, OpenThai and the tiny ones

Within a week of Jev, individuals and small teams had rebuilt its shape from open bases. Some publish better calibration than the original. A guide to the ones worth knowing.

Biplov Gautam

The first System One models with a company behind them get the attention. The reproductions are where the interesting engineering is happening, because their authors had to make every choice in the open and measure it. This is a guide to the ones on the registry, all Apache-2.0 or MIT unless noted.

Bespoke-Nimble-9B — curated data over volume

Bespoke Labs trained a LoRA on Qwen3.5-9B with only 2,676 examples, chosen contrastively, and scores the allowed answer tokens directly for enums, booleans and rubric levels. On their 324-example held-out set it reaches 90.1% (Jev: 93.2%) with an expected calibration error of 0.054; on thirteen human-labelled public subsets, 74.8% macro against Jev's 76.0%. The data, recipe and benchmark suite are all public, and a keyless demo on Modal speaks /v1/systemone at 106 ms median on an H100.

Decider — the most downloaded

Mapika's Decider family (0.8B, 2B, 4B, 35B-A3B and a 2B-vision model) merges Qwen3.5 fine-tunes trained on about 95 public decision datasets, then applies calibration-aware RL and a rank-64 LoRA, and softmaxes letter logits at an answer slot. Decider-2b reports 80.2% on its 67-task regression set with a calibration error of 0.038, and a 3.2 ms median with CUDA graphs. With 130,000 downloads it is the reproduction most people are running.

Von — the encoder route, order-invariant

Von takes the same road as Laya: ModernBERT-large with an option-marker head that scores every option's marker in one bidirectional pass. Version 1.2 made the scoring order-invariant. 63.9% on JevBench's public standard tier with a calibration error of 0.045 (fitted in-sample), 8,192-token context, about 18 ms on a GPU, and an SDK on pip and npm.

JevK5 — distilled from bigger teachers

Alibi Serikbay's JevK5 distils 17,408 teacher questions (from Qwen3.6-27B and GPT-6) plus 30,000 public rows into Qwen3.5 checkpoints at 2B, 4B and 9B, read out as a temperature-scaled softmax over answer letters. Version 0.3 scores 78.4% on JevBench's hard tier with a calibration error of 0.035; the previous version ranked second of 76 entries on the public leaderboard. 13 ms median on an H100.

OpenThai-SystemOne — a language first

iApp's model is Qwen3.5-0.8B continued-pretrained on about 5B Thai tokens, its LM head replaced by a 256-way slot head with an abstain slot, trained on 2–3M decisions and calibrated per question type. 74.3% macro on Bespoke's public subsets, within two points of Nimble at a ninth of the size, and a free hosted endpoint.

The small ones

  • cua-s1-forms: 706,048 parameters and 2.8 MB. A byte-level transformer that decides, per form element, whether to fill, check, click or skip. 100% on a 196-decision real-form evaluation against 83.6% for hosted Jev on the same task. MIT.
  • NanoJev: a 0.6B Qwen3 replica trained on Maze, Snake and ViZDoom, with its full training pipeline. Snake 8/8, ViZDoom Basic 128/128 (Jev 56/128).
  • Jev-Omni: the multimodal one — text, image, audio and video on Gemma 4 12B, 87.6% on its own DecisionBench with 0.040 calibration error. Note the Gemma base's terms before redistributing.
  • The Qwen3.5-4B scorer: one of the first, from 16 September, and non-commercial (CC-BY-NC-4.0) because of its training data.

What they agree on

Every one of these serves TypeSafe's /v1/systemone contract, so a client moves between them by changing a URL, and every one publishes a calibration number — which the original does not. The comparison page puts four of them side by side; each model page names the benchmark behind every number, because these suites do not agree with each other either.