Jev directory
CommunityJev Models & Reimplementations
Intern-Decision-4B
InternLM's multimodal structured decision model, fine-tuned from Qwen/Qwen3.5-4B: a shared state, a schema of named questions and optional images go in, and every question returns an answer distribution from one forward pass. The card reports a 90.02 seven-suite average.
- Reading time
- 2 min
- Published by
- Community
- Author
- InternLM
- Last verified
- Oct 2026
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- Fine-tuned from Qwen/Qwen3.5-4B; accepts state, a schema of named questions and optional images.
- One causal forward pass reads logits before each <decision> placeholder and softmaxes only allowed single-token symbols.
- Self-reported seven-suite average 90.02 vs Jev's 88.74; JevBench-Hard 73.87, Typed Decision 80.55, Brier 0.347, ECE 0.065.
- Latency on one RTX 4090 (local HF path): 44.16 ms mean, 44.03 ms median, 44.60 ms P95 per query.
- Quickstart: Python 3.12+, requirements.txt, then DecisionEngine(device="cuda").predict(request) returns Jev-compatible JSON.
Quickstart
python
from inference import DecisionEngine
engine = DecisionEngine(device="cuda")
request = {
"state": "The customer was charged twice and asks for the extra payment back.",
"questions": {
"team": {"type": "choice", "instructions": "Which team should handle this request?",
"criteria": {"billing": "Payments and refunds", "delivery": "Shipping and delivery"}},
"refund_requested": {"type": "noul", "instructions": "Is the customer asking for a refund?"},
},
}
response = engine.predict(request)
print(response["answers"])Watch out
Card reviewed 2026-10-17; the seven-suite benchmark table, Brier/ECE figures and RTX 4090 latencies are the publisher's own measurements, and the card notes latency is workload and hardware dependent. The calibration pilot is a separate 96-case diagnostic, not an accuracy benchmark, and competitor rows are the publisher's own runs.
Questions about this entry
Short answers drawn from the linked project or page.
InternLM's 4B multimodal structured decision model, fine-tuned from Qwen/Qwen3.5-4B. It accepts a shared state, a schema of named questions and optional images, and returns an answer distribution for every question in one model forward pass.
Options are mapped to single-token symbols (A-Z, a-z, 0-9); the prompt carries one <decision> placeholder per field; logits are read immediately before each placeholder in a causal HF forward pass, softmaxed over the field's allowed candidates and calibrated into typed JSON answers.
Self-reported averages over seven suites: 90.02 average for Intern-Decision-4B versus 88.74 for Jev, with Brier 0.347 and ECE 0.065. JevBench-Hard is 73.87 and Typed Decision 80.55; no gold answers are inserted into prompts.
Per-query end-to-end latency with the local HF inference path: 44.16 ms mean, 44.03 ms median and 44.60 ms P95. The card warns these values are workload and hardware dependent.
A separate 96-case diagnostic using exact reference distributions. The card reports overall Brier/ECE improving from 0.628/0.213 to 0.550/0.089 for the 4B model, which used its separately fitted T=1.992418; the pilot was not used to fit or select the published temperature.
Use Python 3.12+, install requirements.txt in a PyTorch/CUDA environment, then import DecisionEngine and call engine.predict(request): one request dict in, one JSON-serializable Jev-compatible response out. The engine defaults to the checkpoint next to inference.py, and backend="hf" is the only implemented backend.
Yes: requests can carry optional images alongside the state, and the weights are Apache-2.0. The published benchmark and latency tables measure text decisions, and the card reports no separate image-input decision benchmark.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
X: Standard One SH (StandardThinking)
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
InternLM's 0.8B multimodal structured decision model, fine-tuned from Qwen/Qwen3.5-0.8B: a state, named questions and optional images produce an answer distribution per question in one forward pass. Self-reported seven-suite average 79.38, Brier 0.530.
Jev Models & Reimplementations#community#open-models#multimodal
Community
InternLM's 2B multimodal structured decision model, fine-tuned from Qwen/Qwen3.5-2B: a state, named questions and optional images produce an answer distribution per question in one forward pass. Self-reported seven-suite average 84.68, Brier 0.437.
Jev Models & Reimplementations#community#open-models#multimodal
Community
A small distilled sibling of Perplexity's pplx-decider-v1.2: a Qwen3.5-4B decider that reads a state plus typed choice, noul and score questions in one pass, keeping most of the 27B model's Decision Index.
Jev Models & Reimplementations#community#open-models#multimodal
Community
Perplexity's 27B decision model rebuilt on Qwen3.8-27B with a joint request format that answers a whole request's questions in one sequence, reporting 68.09 on Decision Index 0.3 and 77.43 on vision.
Jev Models & Reimplementations#community#open-models#multimodal
Community
A 4B decision model taking text, JSON, or images up to 262,144 tokens: 52.75 on Decision Index 0.2.1, 204 of 231 on JevBench public, and 26.0 ms for a three-question request on one H200.
Jev Models & Reimplementations#community#open-models#local
Community
A Laya fine-tune from Convai Innovations on the 1,200-case typed-decisions benchmark: it reports 0.777 accuracy on the official 400-case test set, above TypeSafe Jev 1.13.0 at 0.727 and the teacher ceiling at 0.735.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.