Published
djev reads DiffusionGemma as a decision API, and sits third on JevBench
Florian S's JevBench table puts djev at 74.3, 1.1 behind Jev. Maisa's djev.dev hosts the API at an announced $0.035 per million input, with native image questions. Matt Mastracci posted phone-camera clips of DiffusionGemma-as-Jev. Davipar/djev-dev is the Apache 2.0 local stack.
Florian S posted JevBench v1.2.6 on September 21 with djev third, at 74.3, and tagged Matt Mastracci. The table we read is now v1.2.8. Jev 1.13.0 is 75.4. SemIf is 74.7. djev is still 74.3.
JevBench credits that hosted row to Maisa (David Villalón) at djev.dev. The llms.txt is the contract: POST https://api.djev.dev/v1/request, request model djev, response djev-0.1. Access is an invitation code, then a recovery key shown once. Up to 32 questions, 20,000 state characters, one state image, and six images across the request. The announced rate is $0.035 per million input tokens, output free. Prepaid billing is still being prepared. The page says probabilities and confidence are experimental and uncalibrated.
Matt posted a phone-camera clip on September 18 of DiffusionGemma-as-Jev doing near-real-time vision. Later posts in that week put a 7x5 grid at about 700 ms per frame and a 5x5 at about 500 ms. Those timings are his clips, not JevBench.
The open stack is Davipar/djev-dev, Apache 2.0. It runs Google’s DiffusionGemma through a patched vLLM (upstream PR 57250), seeds a compact answer canvas, and reads the allowed label IDs after one denoising step. Local clients hit POST /v1/request. The README credits Matthew Mastracci for the structured-generation work this implementation builds on. A 76.87 ms p50 / 86.40 ms p95 figure there is a historical 1,000-request text run on an earlier quantized setup. The notes say that is not a benchmark of the current BF16 release and does not apply to image requests.
On JevBench, djev’s hard-tier accuracy is 69.5% against Jev’s 74.1%. Speed is the high axis, 91.4, at 0.24 s p50 on the production API. Calibration is 65.4 against Jev’s 82.7. We did not create an account or run the model.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Florian's September 21 post lists djev at 74.3, third, and tags @mmastrac. The table we read is JevBench v1.2.8: Jev 1.13.0 75.4, SemIf 74.7, djev 74.3. djev Intelligence 88.4, Calibration 65.4, Speed 91.4, Cost 57.6, hard-tier 69.5% against Jev 74.1%, p50 0.24 s on a production API. Cost uses the announced $0.035 per million input ($0.026 per 1,000 decisions); the note says nothing is charged yet. JevBench credits the hosted row to Maisa (David Villalón) at djev.dev. llms.txt: POST https://api.djev.dev/v1/request, request model djev, response model djev-0.1, invitation to join, up to 32 questions, 20,000 state characters, one state image and six images across the request. Prepaid billing is still being prepared. Probabilities are labelled experimental and uncalibrated. Matt's September 18 clip is DiffusionGemma-as-Jev on a phone camera. Davipar/djev-dev is Apache 2.0, local POST /v1/request on DiffusionGemma plus a patched vLLM (PR 57250), and credits Matthew Mastracci for the structured-generation work. A 76.87 ms p50 figure in that README is an earlier quantized text run, not the BF16 release and not image requests. We did not create an account or run the model. TypeSafe's Master Customer Agreement section 2.3(f) forbids publishing benchmarks of the Services; the JevBench row is reported as published.
Compare
SemIf is the closest open row (74.7, 0.7 behind Jev) and reads option logits from Qwen3.5-4B. djev is a hosted DiffusionGemma API with native image input, which Jev's models page does not offer. SimpleJev, covered separately, reads next-token logits from stock open models and trails on the composite (Qwen3.8-27B 67.3) while matching Jev's 75.0% hard-tier accuracy on that row. jeff and Kev copy POST /v1/systemone. djev's local path is POST /v1/request. Calibration 65.4 is the weak axis next to Jev 82.7.
Terms
- djev
- Maisa's hosted decision API on DiffusionGemma, and the name Matt Mastracci used for DiffusionGemma-as-Jev. JevBench v1.2.8 lists the hosted row at 74.3.
- Structured read
- djev's inference path: seed a compact answer canvas, denoise once, and read probabilities for the allowed label token IDs instead of generating text.