Most large models are tuned to generate text; JEV-35B is tuned specifically to make rapid, calibrated decisions from mixed text and image inputs. That focus makes it useful wherever you need one-shot ranking, routing or binary decisions at scale rather than conversational generation.
What Sets It Apart
- Decision-first design: reads a state, a single question, and up to 256 enumerated options in one pass and returns a calibrated probability distribution, a chosen option, or a 0–5 score. This makes it suitable for routing, short classification workflows, UI element selection and other discrete-choice tasks.
- System 1 (fast, no chain-of-thought): the model is optimized for single-step decisions with low latency (median ~204 ms per /v1/decide request on a B200 in the authors' tests). It exposes a dedicated POST /v1/decide API and includes a merged decision head and calibration parameters to produce reliable probabilities.
- Practical deployment options: bf16 weights (~72 GB download, ~66.5 GiB GPU memory) and an NVFP4 quantized variant (~25.6 GB download, ~23.3 GiB GPU memory) let teams run the model on large GPU servers or a single 32–48 GB GPU with minimal changes to the provided serve.sh/vLLM setup.
Key Capabilities
- Choice decisions (2–256 options), yes/no (noul) and 0–5 scoring supported out of the box; can read mixed lists of text and images as the decision state. Calibration and per-kind temperatures are included in model artifacts to improve probability outputs.
- Multimodal zero-shot image decision ability via the Qwen3.5-35B-A3B vision tower; authors report an approximate public vision score ~71.9 on a rebuild of Decision Index vision benchmarks.
- Benchmarked as a System 1 decision model with a reported Decision Index 0.3 public value ~59.73 (authors' scoring kit); shows strong retrieval/classification and tools/automation skills in the reported suite.
Who it's for — and trade-offs
Great fit if you need a dedicated, low-latency decision/ranking component that returns calibrated probabilities for downstream routing, automation pipelines, or UI control. The NVFP4 build is a practical option for single-GPU hosting. Look elsewhere if you need a conversational/chat model, heavy multi-step reasoning, or best-in-class numeric reasoning: JEV-35B is explicitly a System 1 decision model (no chat, no chain-of-thought) and its Knowledge & Reasoning benchmarks lag some peers. Also English-centric and not recommended for high-stakes decisions without additional confidence gating or secondary checks.
