decider-4b
43.1#129 of 261 on the open-weights board · Capability #105
Model guide · Open System One rebuild · Qwen3.5-4B-Base · pip install decider-ai · updated October 26, 2026
Millisecond decisions from 8.4 GB of open weights, on TypeSafe’s wire format.
decider-4b reads a state and a set of typed questions and returns one probability distribution per question from a single forward pass of an open Qwen3.5-4B model. Its server speaks TypeSafe’s /v1/systemone wire format, so TypeSafe’s own SDKs work when pointed at it. It is Apache-2.0, installs from PyPI, and was the first system to finish above Jev on JevBench when it was added; on the re-measured v1.6 pool it is #129.
decider-4b is a third-party project by Mapika. Jev AI does not host it; this page is a guide to what it is and how to run it yourself.
Try it on this page
Edit the example or paste your own text, then run it. Sign in to run: new accounts get 20 welcome credits, and a failed run costs nothing.
Jev AI does not host decider-4b, so this editor runs Jev 1.14 on the same kind of decision. Compare the answers with your own decider-4b setup.
Support ticket triage · Jev 1.14
1 Input
Hi, this is the third time I am writing. Our payouts have failed every day since Monday and my team cannot pay suppliers. I already re-entered the bank details twice. If this is not fixed today we will have to move to another provider.
2 Questions
One benchmark measured every system under one method. It ranks 261 open-weights systems on one board and 54 hosted APIs on another; Jev, a hosted API, is ranked on the second and shown on the first as an unranked reference. The full board is on the JevBench results page.
decider-4b
43.1#129 of 261 on the open-weights board · Capability #105
Jev 1.13.0
71.4Open-weights reference row · #18 of 54 on the API board · Capability #15
IntelligenceHow often it picks the right answer, half from sealed decisions
CalibrationWhether 0.8 really means about 80%
SpeedMeasured response time
Open half 45.2 · sealed half 35.0; chance-corrected by request type: choice 65.0, yes/no 19.1, score 36.3. The composite also weighs a fourth axis, cost, which these bars leave out. Self-hosted latency is adjusted by the benchmark to approximate production load.
decider is a family of open models from Mapika that reproduce the System One model class without generating text. The 4B model is Qwen3.5-4B-Base fine-tuned to read a state and typed questions and return one probability distribution per question in one pass. Its README states that it is not affiliated with TypeSafe and that nothing was distilled from Jev: the training mixture is public datasets plus data labelled by a local Qwen3.5-27B teacher.
The server implements TypeSafe’s /v1/systemone wire format and the README says TypeSafe’s SDKs work unchanged with a changed base URL. The 4B needs one CUDA GPU with about 8.4 GB in bf16 and captures its CUDA graphs at start-up; the package also runs on CPU, and Apple Silicon support covers the smaller dense models.
JevBench records one disclosure from the author: 8,000 rows in the v2 training stage came from generators written from the published names of the ten sealed-question families, without reading any sealed item. The benchmark’s external review judged the result legitimate.
JevBench v1.7.48 ranks decider-4b v2 #129 of 261 open-weights systems at 43.1, measured on an RTX 5090 at a 25ms raw median per decision. Speed is 90.3, below Jev’s 91.2, and calibration 89.6 against Jev’s 90.6. Intelligence is 40.1 against Jev’s 63.6, and 35.0 on the sealed half against 61.6; its open-minus-sealed gap of 10.2 points is about four times the field median. Because intelligence sits below the benchmark’s floor of 50, the composite is scaled down, which is why the score is 43.1 while the individual axes look respectable. On the Capability ranking, which leaves speed and cost out, it is #105.
Mapika publishes several decider models and releases come quickly. JevBench v1.7.48 ranks five of them; the 12B pair added on 8 October are the strongest.
decider-4b v2by Mapika
Qwen3.5-4B-Base, 4.2B parameters, about 8.4 GB in bf16. Released on 24 September; the version JevBench measured, kept under the Hugging Face tag v2.
#129 of 261 at 43.1
decider-12b v2by Mapika
Gemma 4 12B fine-tuned the same way, added to the board on 8 October: intelligence 63.0, close to Jev’s 63.6, with a weaker calibration axis.
#36 of 261 at 73.7
decider-12b v1by Mapika
The first 12B release, read out of a stock Gemma 4 12B, added on the same day.
#43 of 261 at 73.2
decider-4b v2.1by Mapika
The current default, released the same day: v2 plus a further LoRA stage. Its author reports it slightly weaker on JevBench’s public hard items, 0.649 against 0.676, and less well calibrated on hard items.
Not measured in JevBench v1.7.48
decider-35b-a3bby Mapika
Qwen3.5-35B-A3B-Base, 34.7B parameters with 3B active: about 65 GB in bf16 or 19.6 GB in NVFP4.
#96 of 261 at 66.5, held back by intelligence
decider-2bby Mapika
Qwen3.5-2B-Base, 1.9B parameters, about 4 GB. The first decider model.
#205 of 261 at 8.0, held back by intelligence
decider-0.8b and decider-2b-visionby Mapika
A smaller text model and a vision-language variant that reads decisions from pixels, still on older text weights.
Not measured in JevBench v1.7.48
Install the package from PyPI, start the server on a CUDA GPU, and point a TypeSafe SDK or plain HTTP client at it. Pin the Hugging Face tag v2 if you want the exact build JevBench ranked.
pip install decider-ai
# Start the server per the README; the 4B weights download from huggingface.co/Mapika/decider-4b
# (tag v2 is the build JevBench measured; v2.1 is the current default).
curl http://localhost:8000/v1/systemone \
-H "Content-Type: application/json" \
-d '{"state": "Review: The app crashes every time I open the camera tab.",
"questions": {"sentiment": {"type": "score", "instructions": "How negative is this review?",
"criteria": ["neutral", "mildly negative", "very negative"]}}}'From the project’s own documentation and JevBench v1.7.48.
decider-4b is close enough, and fast enough, that the honest test is your own cases. Run them on Jev here with no setup and keep the answers: if decider-4b agrees where it matters, you have a strong case for self-hosting, and if it does not, you will see exactly which cases it misses.
An Apache-2.0 open model from Mapika that reproduces the System One model class on Qwen3.5-4B-Base: it reads a state and typed questions and returns a probability distribution per question in one forward pass, behind TypeSafe’s /v1/systemone wire format. It installs from PyPI as decider-ai.
No. Its README states that it is not affiliated with or endorsed by TypeSafe AI and that nothing was distilled from Jev.
decider-4b v2, released on 24 September 2026 and available under the Hugging Face tag v2. The current default, v2.1, was not measured on the v1.6.1 pool; its author reports it slightly weaker on JevBench’s public hard items.
Yes, according to its README: the server implements POST /v1/systemone in TypeSafe’s wire format, so the SDKs work once their base URL points at your server. Check input lengths against the 32k-token context first.
One CUDA GPU with about 8.4 GB free in bf16. The package also runs on CPU, and Apple Silicon support covers the smaller dense models. JevBench measured it on an RTX 5090.
Every figure on this page is quoted from a published source and was read on October 26, 2026.
decider-4b and every other product named on this page belongs to its respective owner and is not affiliated with Jev AI. Figures are quoted from the sources above as they read on October 26, 2026 and can change.