Jev directory
CommunityJev Models & Reimplementations
Intern-Decision-2B
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.
- 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/Qwen3.5-2B; each request carries a state, named questions and optional images.
- Structured candidate scoring: no generate() call, no sampled free-form text, one forward pass per request.
- Self-reported seven-suite average 84.68 vs Jev's 88.74; JevBench-Easy 100.00, Hard 63.96, Brier 0.437, ECE 0.100.
- RTX 4090 latency: 33.28 ms mean, 33.15 ms median and 33.55 ms P95 per query, end to end.
- Supports 1-16 questions per request with up to 62 options each; inference module ships with the checkpoint.
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 benchmark table, Brier/ECE values and RTX 4090 latency figures are the publisher's own measurements and the card warns latency is workload and hardware dependent. This card reports no calibration pilot, so calibration is described only by Brier 0.437 and ECE 0.100.
Questions about this entry
Short answers drawn from the linked project or page.
InternLM's 2B multimodal structured decision model, fine-tuned from Qwen/Qwen3.5-2B. It accepts a shared state, a schema of named questions and optional images, and returns an answer distribution for every question in one forward pass.
It performs structured candidate scoring: options map to single-token symbols, the assistant skeleton carries one <decision> placeholder per field, and logits are read immediately before each placeholder, softmaxed over allowed candidates and calibrated.
Self-reported seven-suite average 84.68 versus Jev's 88.74, with Brier 0.437 and ECE 0.100. JevBench-Easy 100.00, Jevbench-Original 84.72, Jevbench-Hard 63.96, Typed Decision 79.35, ToolACE 96.45 and AG News 89.96.
Per-query end-to-end latency with the local HF inference path: 33.28 ms mean, 33.15 ms median and 33.55 ms P95. The card warns the values are workload and hardware dependent.
It is the middle size of the Intern-Decision family, using the same request schema and calibration approach. The card's seven-suite averages are 79.38 for the 0.8B, 84.68 here and 90.02 for the 4B model.
Use Python 3.12+, install requirements.txt in a PyTorch/CUDA environment, then import DecisionEngine and call engine.predict(request); the engine defaults to the checkpoint next to inference.py and returns a Jev-compatible JSON response.
Apache-2.0. The card's benchmark table and latency figures are the publisher's own runs, no calibration pilot is published for this size, and image input is supported while the reported results measure text decisions.
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 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.
Jev Models & Reimplementations#community#open-models#multimodal
Community
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
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.