A 395M-parameter reflex engine that plays Doom without a policy net

What it does

Von is a 395M-parameter bidirectional ModernBERT model that performs discrete decisions—classification, routing, guardrails, and ordinal scoring—in a single non-autoregressive forward pass. Instead of decoding text token-by-token, it scores options in parallel against a premise and emits calibrated probabilities and confidence intervals. It runs entirely offline across CUDA, ROCm, Metal, OpenVINO, or CPU and claims full compatibility with the TypeSafe System One protocol.

The interesting bit

The architecture deliberately prevents options from attending to each other, guaranteeing that shuffling the answer order cannot change the result—a flaw the authors note plagued v1.1 and most rival option-packing models. That same fixed-weight model can answer routing questions or, wired to a depth-buffer text renderer, control Doom zero-shot with no policy network or reinforcement-learning fine-tuning.

Key highlights

• Order-invariant option scoring via masked cross-option attention; v1.2 scores 0% answer flips under reordering on JevBench hard tier, down from 49.5% in v1.1.
• Sub-18 ms local inference on a GPU versus ~115 ms for the proprietary TypeSafe Jev API.
• Calibrated uncertainty via joint Cross-Entropy and Brier loss, with a fitted temperature of 1.1692.
• Three decision primitives: decide (categorical), judge (binary probability), and rate (ordinal expectation).
• Outperforms Jev on zero-shot ViZDoom combat (9.00 vs. 5.62 kills) and runs on Intel iGPUs via OpenVINO.

Caveats

• Still trails the proprietary TypeSafe Jev on the 49-task jabr v2 language-and-logic suite by a wide margin (72.0% vs. 96.6% macro accuracy).
• JevBench hard-tier accuracy sits at 38.7%, unchanged from v1.1, and the standard tier dipped slightly from 65.3% to 63.9%.

Verdict

Worth a look if you need deterministic, low-latency routing or guardrails on commodity hardware and are willing to trade generative flexibility for speed. If your use case demands state-of-the-art broad-domain reasoning or you are already satisfied with a cloud-hosted proprietary decision API, this is probably not your upgrade.