The open side of the decision-model class was already moving when this month's arXiv wave landed, and the wave added training recipes, architectures, and language ports faster than any leaderboard can absorb. Here is the shape of it: what the open models are, how they are trained, and what to actually pick by.
The training recipes
OpenJev-RLCD is a working implementation of the RLCD idea TypeSafe described: the model samples a rationale, then a strictly proper scoring rule grades the answer distribution it commits to. With Qwen3-1.7B it matches or beats SFT, RFT/STaR, and GRPO in accuracy and beats all of them on selective prediction. FunDM skips training entirely, reading decisions from frozen language models; Jeva builds the machinery from first principles. Dyad extends an LLM with native typed decision-making, and PACT tunes single-token decisions with pairwise anchoring. LLM-as-Jev argues fine-tuning is mostly unnecessary — the LLM is already a decision model when you read it correctly.
Architectures beyond the encoder
SanSi loops a typed decision model for System 1.5 thinking; JevCascade studies readout stability and compute allocation for prefill-only decisions; FastJEV attacks redundancy for compact inference; NumericJev decodes numbers with multiway decision trees. On the serving side, LAVOIR puts the decision and a value-of-information question in one forward pass, so the model can ask for missing slots instead of guessing.
The ports and the languages
Chinese-Jev brings the shape to Chinese-language tasks, and the Hugging Face ecosystem now carries the ports: Laya multilingual routes by script, OpenThai-SystemOne is a Thai build with GGUF, Ollama, and CoreAI ports, and StartLux-Decision ships a family from 0.8B to 35B with quantizations. Turkish, Japanese, and biomedical variants are all in the catalog, alongside Laya typed-decisions GGUF and ONNX ports for runtimes without PyTorch.
What the open models give you
- A same-shape API. Most serve
POST /v1/systemoneor a Python API that mirrors the SDK, so the switch is a base URL or a model field. - Weights you can patch. The labels-override-definitions study found one open code base immune to option-label bias by rendering only the definition — a fix you can make yourself in open weights, and one that the hosted model comparison could not apply.
- Measured baselines. Laya's reproduction reproduces the card's 0.766 and finds it uniformly under-confident (ECE 0.214), which a temperature fit drops to 0.037.
- Calibration you control. The distilled student study shows sentence encoders can carry typed decisions in-domain, which is what makes a small local model viable.
Pick by runtime and calibration, not the leaderboard
The student-guided routing paper is the honest summary of this space: open checkpoints compete on cost, latency, and control, while accuracy depends on question design and the candidate set you actually run. Benchmark the port you can serve — GGUF on a laptop, MLX on a Mac, ONNX in the browser — on your own rows, and measure calibration and rejection before accuracy. The open weights article covers the hub side, and choosing an open decision model covers the decision itself.
