Quick start
pip install decido-ai # macOS (Apple Silicon) and Linux x86_64: prebuilt, nothing to compile
from decido import Decido m = Decido("decido-ai/decido-1-26b-a4b") # downloads once from the Hub, then runs locallyprint(m.decide( "Customer: I was charged twice for order 4821. Please refund one of the charges today.", { "refund": "Is the customer asking for a refund?", # yes / no "team": {"q": "Which team should handle this?", "options": ["billing", "shipping", "tech"]}, "urgency": {"q": "How urgent is it?", "scale": ["low", "normal", "urgent"]}, },))# prints something like: {"refund": {"yes": 0.99, "no": 0.01}, "team": {"billing": 0.99, "shipping": 0.01, "tech": 0.0}, "urgency": {"low": 0.12, "normal": 0.44, "urgent": 0.44}}
Command line and HTTP:
decido run decido-ai/decido-1-26b-a4b "Customer: charged twice, refund please" -q refund="Is a refund requested?" -q team="Which team?|billing,shipping,tech"decido serve decido-ai/decido-1-26b-a4b --port 8080 # Jev-compatible API: POST /v1/systemone
Existing Jev clients work unchanged against decido serve (point the SDK at it: TYPESAFE_BASE_URL=http://127.0.0.1:8080).
Question formats. "text" → yes/no · {"q": ..., "options": [...]} or {"q": ..., "options": {"key": "description"}} →
pick one · {"q": ..., "scale": [...]} → ordered scale. m.predict({"state": ..., "questions": ...}) takes the full Jev 1.13 wire
format (noul / choice with 2–255 options / score with 2–10 levels) and returns Jev-style answers (noul, choice +
probabilities + confidence, score + legend). The document may be text or JSON; it is treated as untrusted data
(instructions inside it are not followed; control-token strings are neutralised).
Mac GPU: on Apple Silicon decido (≥ 1.0.3) runs the GGUF on the Mac GPU (Metal) by default — 1.1 s on an M4 Pro (2–4× faster than CPU). Metal agrees with the CPU file on 98.5–99.5% of benchmark questions and is deterministic across calls; device="cpu" reproduces the published CPU numbers bit for bit.
On a GPU (pip install "decido-ai[gpu]", Python ≥ 3.10), Decido("decido-ai/decido-1-26b-a4b", device="cuda") runs the bf16 weights in this repo
with transformers. Plain transformers, no package: see example_transformers.py in this repo.
The Decido 1 family
Table with columns: model, built on, runs on, six-group score (Jev 1.13 = 83.85), latency, M4 Pro: CPU · Mac GPU (1k-token document, 5 questions)| model | built on | runs on | six-group score (Jev 1.13 = 83.85) | latency, M4 Pro: CPU · Mac GPU (1k-token document, 5 questions) |
|---|
| Decido-1 (this model) | Gemma-4-26B-A4B | laptop CPU (~30 GB RAM) · GPU | 85.38 (CPU) | 4.1 s · 1.1 s |
| Decido-1-Fast | Qwen3-1.7B | laptop CPU (~8 GB RAM) | 73.01 (CPU) | 0.64 s · 0.30 s |
| Decido-1-Ultra | Qwen3-0.6B | laptop CPU (~4 GB RAM) | 62.16 (CPU) | 0.41 s · 0.13 s |
| Decido-1-Max | Gemma-4-26B-A4B | laptop (slow) · GPU path coming | 89.23 (GPU) | ~28 s / question · ~11 s / question |
| Decido-1-Think | Gemma-4-26B-A4B | laptop (slow) · GPU | 88.71 (GPU) | ~19 s / question · ~10 s / question |
| Decido-1-XL | DeepSeek-V4-Flash | 4× A100 80 GB | 87.25 (GPU) | GPU only (~0.45 A100-s / question) |
Results
Six benchmark groups used by the open Jev-reproduction community (JevBench, Kev, OpenJev text tasks, Nimble, VitaminC,
MASSIVE-en; frozen copies, same harness for every model). Score = accuracy-type metric per group, then the mean.
Table with columns: benchmark group, Decido-1, CPU (Q8_0 GGUF), Decido-1, GPU (bf16), Jev 1.13 (published)| benchmark group | Decido-1, CPU (Q8_0 GGUF) | Decido-1, GPU (bf16) | Jev 1.13 (published) |
|---|
| JevBench | 87.61 | 88.17 | 87.18 |
| Kev | 84.35 | 84.46 | 85.52 |
| OpenJev (18 tasks) | 78.37 | 78.39 | 72.96 |
| Nimble | 87.23 | 88.65 | 91.84 |
| VitaminC | 86.14 | 87.15 | 78.46 |
| MASSIVE | 88.57 | 88.57 | 87.14 |
| six-group mean | 85.38 | 85.90 | 83.85 |
CPU: five groups measured CPU-only; OpenJev on the same GGUF via llama.cpp CUDA (all 222,054 requests; a CPU-only 300-per-task
sample agrees on 97.4% of items). GPU bf16 95% bootstrap CI: 84.70–87.13.
Calibration (multi-class Brier and 10-bin ECE; lower is better; ✓ = at or better than Jev):
Table with columns: check (CPU, shipped GGUF), Brier (lower is better), Jev 1.13, ECE, Jev 1.13| check (CPU, shipped GGUF) | Brier (lower is better) | Jev 1.13 | ECE | Jev 1.13 |
|---|
| JevBench | 0.169 ✓ | 0.181 | 0.038 | 0.032 |
| Kev v4 dev | 0.210 ✓ | 0.211 | 0.029 ✓ | 0.049 |
| Kev v9 dev | 0.267 | 0.212 | 0.020 ✓ | 0.034 |
| VitaminC | 0.202 ✓ | 0.313 | 0.041 ✓ | 0.104 |
| MASSIVE | 0.164 ✓ | 0.204 | 0.025 ✓ | 0.075 |
Brier ≤ Jev on 4/5 checks, ECE ≤ Jev on 4/5. Robustness: identical outputs across repeated calls, request order and
thread count on CPU (0/50 requests changed); option-order flip rate 4.8% pooled (1.4% on GPU with 4-order averaging, perm=4);
compound questions: 96.7% accuracy asked together vs 97.5% asked separately.
Speed (Apple M4 Pro, CPU only, 10 threads, ~1k-token document + 5 questions): p50 4.14 s, p95 4.26 s; one-time load ~97 s;
~30 GB RAM. The document is read once and cached; each extra question adds a few hundred ms.
How it works
- Prompt: system instruction + document in
<document> + one question block per question; questions are answered
independently from the cached document.
- Readout: the next-token probability of single-token labels (Yes/No; A–Z, or A0–Z9 above 26 options; 0–9 for scales),
normalised over the labels, then a per-type temperature fitted on held-out tasks (noul 1.027, choice 1.146, score 1.013).
- Files:
decido-1-26b-a4b-q8_0.gguf (CPU, 26.9 GB), bf16 safetensors (GPU, 48 GB), lora/ (the adapter alone, for
google/gemma-4-26B-A4B-it), bundle/ (frozen prompt template, tokenizer, temperatures), decido.json (manifest).
- Training: LoRA r16 (attention, dense MLP, router, all 128 experts), soft cross-entropy + Brier + ranked-probability loss on
~163k questions (gold labels 0.7 + temperature-calibrated Qwen3.8-27B soft labels 0.3), 2 rounds; round chosen by a
pre-registered dev rule.
Disclosures
- In-domain supervised. Training used the train splits of several datasets whose test/eval splits are in the benchmark
groups (e.g. MASSIVE, VitaminC, Open-Jev tasks). Decido-1's recipe without any train split of a scored dataset scores 82.04
(below Jev's 83.85), so the margin over Jev should be read as in-domain supervised. Jev was evaluated zero-shot.
- Jev numbers are AutoTrust's published run of Jev 1.13 (run-to-run spread up to 1.6 points), not Jev run through our
harness. OpenJev is scored on 18 of its 19 tasks (the GSM8K-rerank pool is unpublished); Jev's 72.96 is a 19-task mean.
- No Jev outputs and no Claude outputs were used for training; no benchmark test split was trained on (13-gram + exact-match
leakage gates).
Limitations
- English only; trained for document-grounded decisions (classification, policy checks, entailment, routing, triage, rating).
It is not a chat model and not a general reasoner — use it with the
decido package (plain chat use does not give calibrated
probabilities).
- Weakest on hard knowledge multiple choice (Kev v9: calibration worse than Jev) and multi-step arithmetic.
- Probabilities are calibrated on held-out tasks, not on your data — check calibration on a labelled sample before relying on
thresholds. Do not use it as the only basis for decisions with legal, medical, financial or safety consequences.
License
Apache-2.0 (weights and code). Base model: Gemma-4-26B-A4B-it (Apache-2.0). Training-data sources and their licences: NOTICE.