Design and compatibility

OpenJev Multimodal independently implements the Jev-style SystemOne interface and adds local image input. It is not affiliated with TypeSafe and does not contain proprietary Jev weights.

Compatible interface

The request envelope, Noul/Choice/Score types, named answers and probability fields follow the public interface. jev-latest selects your local Qwen model. Structured instructions and criteria are supported.

One synthetic request containing all three types was sent to the official online API using the local typesafe-ai skill. HTTP 200 from jev-1.13.0 confirmed the wire shape. This was not an accuracy or speed comparison.

Predictions, local limits, confidence and usage accounting differ. We support 2–255 Choice options, up to 64 Score levels and 64 questions. TypeSafe currently documents up to 10 Score levels. See /v1/limits for actual local limits.

Probability readout

1. Validate and sanitize state and images.
2. Apply the backend's native chat template with thinking disabled.
3. Map options to verified single-token labels.
4. Add the same +100 logit bias to each label. Request one token and post-sampling probabilities, without top-k/top-p/min-p truncation or repetition penalties.
5. Require every candidate in the returned list and normalize over those candidates. Missing values cause a 502.
6. Compute typed answers directly from the distribution.

For selected labels, a shared bias cancels:

exp(logit_i + b) / sum_selected exp(logit_j + b)
  = exp(logit_i) / sum_selected exp(logit_j)

This holds up to floating-point rounding. Bias brings candidates into the returned list while preserving relative odds. Completeness is checked; missing-label probabilities are never invented. Grammar-only readout was unsuitable because the verified llama.cpp build can expose probabilities before lazy grammar filtering.

Choice returns the argmax. Score is sum(index * probability). Confidence is 1 − normalized entropy. Wording and option order affect predictions; these are not calibrated correctness guarantees.

Latency path

Every question costs a prefill of its full prompt. Qwen3.5 and Qwen3.6 are hybrid models: their recurrent layers cannot roll back to an arbitrary cached position, so llama.cpp re-read the shared state, images included, for every question. Requests with two or more questions now evaluate the shared prefix once; llama.cpp checkpoints it, and each question resumes from there with only its own text. A screenshot with four questions takes less than half the time. Decisions stay the same; probabilities move by at most 0.067 because the prefix runs as its own batch.

The API also fills a verified chat-template skeleton locally instead of calling /apply-template (about 9 ms) on every request, and resizes an oversized image once, straight to the vision encoder's size. Measurements →

Vision path

Bounded images are EXIF-transposed, converted to RGB and re-encoded; images larger than the vision budget are resized once to the backend's own target size (a 32-pixel grid within --image-tokens). The API reads llama.cpp's actual media marker from /props and passes image data through its vision projector. This is native visual inference, not an OCR-only substitute.

Remote fetching and request-controlled file access are disabled. Treat model judgments as fallible application inputs: prompt injection can still affect predictions.

References

Project	Approach
TypeSafe Jev	Proprietary hosted SystemOne
openjev-sglang	Text-only one-token Qwen readout with SGLang
AlexWortega/openjev	Qwen3.5 fine-tuned as a three-way NLI cross-encoder, including a vision checkpoint
This project	llama.cpp/Metal, text + images, one-token typed probabilities

Published NLI model scores do not describe this implementation. Our chart uses our own Qwen3.6 API measurements. The Hacker News discussion motivated this project.