TensorSharp Adds Image-Aware Local Jev API With 3.3x Faster Structured Reads, Zero Schema Failures
fuzhongkai · reddit · 2026-09-24
The TensorSharp maintainer shipped a major update: a local implementation of Jev's core decision API (noul/choice/score) extended to image analysis via an optional images field — up to 8 inline images per request, with image embeddings fed directly into the structured decision path, no 'describe then classify' step.
Method: Inspired by vLLM PR #57250, it uses a one-step structured read — after prompt prefill, label logits are read from an answer canvas and serialized to JSON server-side, so the model never generates probability JSON token by token; questions sharing a canvas share a forward pass.
Results (12 cases × 3 reps vs LocalJev's generated JSON on the same backend): 36/36 vs 27/36 valid requests, 0 vs 9 schema failures, p50 latency 2.877s vs 10.479s, median latency ratio 3.345× across 27 matched successes. Correct decisions were 108/108 on both when valid. Average input length 192.7 vs 589.6 tokens. Text-only results; no image speedup claims. Runs locally with DiffusionGemma Q4KM GGUF plus a 2.8GB vision shard.
More from coding & agent
- 30 lines of JavaScript, no image models: pure-code generative art demo — nc_frey · 2026-09-24
- Claude One-Shots a Full Music Video: P(doom) MV Source Code Goes Open Source — trq212 · 2026-09-24
- Developer calls Claude Opus 5.5 'a doof' at database tasks — rickasaurus · 2026-09-24
- TRACES: A New Benchmark That Grades AI Problem-Solving Process, Not Just Correct Answers — dr_cintas · 2026-09-24
- Dev reverse-engineers Qwen Image 2.1 PE, ships ComfyUI node that auto-computes dimensions — BleynSpecnaz · 2026-09-24
- Cua Releases Cua-S1-4B, First Multimodal Decision Model RL-Trained on Live Computer-Use Tasks — multimodalart · 2026-09-24