sgyaqing/verdict released v0.3.0 in commit ec3452b on 2026-10-26. Verdict is a high-performance serving engine for the System One decision model, and this version handles multimodal for the first time. The new family verdict-models/OneJev1.6 comes in four sizes: 0.8b, 4b, 9b, and 27b. Images and video are handled at /v1/systemone using TypeSafe's assets spec, and images at /v1/decisions using the OpenAI-compatible input_image. Weights are uploaded to both Hugging Face and ModelScope, and the changelog states that the answers were verified against the official model.
The way models are chosen has changed as well. A single repository holds multiple sizes, the size is specified with a :variant tag, and the package can be composed from a manifest alone. On first serving, the entire official repository is snapshotted, including weights and tokenizer, and it is checked for missing items. In the inference path, text and media decoding run inside CUDA graphs, and media that shares prefixes is handled by forked graphs. The torch.compile cache survives restarts.
The request handling rules were also reworked. --max-body-mb was added with a default of 64 MiB, and if --max-input-tokens is not specified, input is truncated at the max_len declared by the model. When the executor process dies, the server stops with a non-zero exit code instead of repeating 500 responses. Deployment is via two Docker Hub images, sgyaqing/verdict:0.3.0-cpu and 0.3.0-cuda (linux/amd64). For performance, the README newly adds throughput and P99 latency graphs for OneJev1.6-9b text, image, and video processing on an RTX 4090, and states up to 66x the throughput of the model maker's server.
The verifiable scope is narrow. The commit stats show 4104 lines added and 2209 lines deleted, and the largest change was executor/torch_exec.py (1334 lines added). The Apache-2.0 license and the NOTICE were included in the image builds, and the NOTICE states that the scheduling design references vLLM, the batching queue design references Text Embeddings Inference, and the graph runner was taken from the mm_engine.py of qev(OmniJev). The multimodal family is only OneJev1.6, and media processing requires pillow 13.0.0 and torchvision 0.29.1. Dockerfile.base-media states that the Qwen video processor requires torchvision even for text-only serving of VL checkpoints. Since no externally verified figures or schedules for supporting other families are presented, this release stays within the range where the System One family can be tried directly.