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-think-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": 0.05, "normal": 0.51, "urgent": 0.44}}
Command line and HTTP:
decido run decido-ai/decido-1-think-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-think-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).
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 | 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 (this model) | 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-Think, GPU, Decido-1, GPU, Jev 1.13 (published)| benchmark group | Decido-1-Think, GPU | Decido-1, GPU | Jev 1.13 (published) |
|---|
| JevBench | 94.33 | 88.17 | 87.18 |
| Kev | 85.63 | 84.46 | 85.52 |
| OpenJev (18 tasks) | 83.63 | 78.39 | 72.96 |
| Nimble | 92.55 | 88.65 | 91.84 |
| VitaminC | 86.98 | 87.15 | 78.46 |
| MASSIVE | 89.14 | 88.57 | 87.14 |
| six-group mean | 88.71 | 85.90 | 83.85 |
Calibration (GPU; ✓ = at or better than Jev):
Table with columns: check (GPU), Brier (lower is better), Jev 1.13, ECE, Jev 1.13| check (GPU) | Brier (lower is better) | Jev 1.13 | ECE | Jev 1.13 |
|---|
| JevBench | 0.098 ✓ | 0.181 | 0.031 ✓ | 0.032 |
| Kev v4 dev | 0.207 ✓ | 0.211 | 0.052 | 0.049 |
| Kev v9 dev | 0.223 | 0.212 | 0.042 | 0.034 |
| VitaminC | 0.205 ✓ | 0.313 | 0.037 ✓ | 0.104 |
| MASSIVE | 0.167 ✓ | 0.204 | 0.053 ✓ | 0.075 |
Brier ≤ Jev on 4/5, ECE ≤ Jev on 3/5. Option-order flips 4.3% pooled (Jev 14.6%).
How these numbers were produced: Think ran as a component of our research system (a gated mixture of Decido-1, Think and Qwen3.8-27B scoring 89.57, later distilled into the single model Decido-1-Max;
vLLM, bf16, one A100); its recorded outputs were re-scored on their own through the same harness (post-hoc; the model was not
selected on these numbers). 157 Kev questions (2.3% of Kev) were never sent to Think by the research system's router and use Decido-1's
answer. Reasoning hit the 1,024-token cap on 46% of OpenJev questions (the answer is then read after a short "answer now" note).
On a laptop CPU (Q8_0 GGUF via decido, Apple M4 Pro, 10 threads): median 19 s per
question (mean 27 s; ~8 s on short classification tasks, up to ~49 s on multi-step ones) — the
reasoning is generated token by token. On 60 benchmark questions (10 per group) the CPU file chose the same answer as the
recorded GPU run 98.3% of the time (greedy reasoning can diverge slightly between bf16 GPU and 8-bit CPU
numerics); a CPU-only accuracy run of the full suite was not done.
Mac GPU: on Apple Silicon decido (≥ 1.0.3) runs the GGUF on the Mac GPU (Metal) by default — ~10 s / question on an M4 Pro (~2× faster than CPU), deterministic across calls. On 20 recorded benchmark questions the Metal run chose the same answer as the GPU run 20/20.
GPU cost: ~0.25 A100-seconds per question in batch (vLLM).
How it works
- Training: from Decido-1's adapter, the model generated its own rationales on ~40k training questions; it was then trained to
give the gold-anchored answer distribution after all of its rationales, including wrong ones ("hybrid STaR"). Training
only on correct rationales made it worse; plain thinking without this step overrode in-domain label conventions.
- Inference: Gemma-4 thinking channel, greedy decoding, cap 1,024 tokens; the label distribution is read at the first token
after the end of thinking. No temperature scaling (T = 1).
- Files:
decido-1-think-26b-a4b-q8_0.gguf (CPU), bf16 safetensors (GPU), lora/ (adapter for google/gemma-4-26B-A4B-it), bundle/,
decido.json.
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).
- Think's scores are post-hoc component scores from the research-system run (see above), not a separately pre-registered evaluation.
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.
- Latency is dominated by the generated reasoning (hundreds of tokens per question); prefer Decido-1 for interactive CPU use.
License
Apache-2.0. Base model: Gemma-4-26B-A4B-it (Apache-2.0). Training-data sources: NOTICE.