Decido
Fast, accurate decision models for AI agents. Runs on your laptop.
Ask a yes/no or multiple-choice question about any text. Get the answer and how sure it is. Open models, more accurate than Jev 1.13, the leading decision API.
pip install decido-ai
Ahead of the leading decision API, without a GPU.
Decido-1 is a mixture-of-experts model with 4B active parameters, quantized to 8 bits. It scores higher than Jev 1.13 on four of the six benchmarks and on the mean. Decido-1-Max, which reasons before it answers, reaches 89.2 and is ahead on all six, in one set of open weights.
Per-benchmark results →Five questions in 0.13 seconds, on a laptop.
Decido-1, Fast and Ultra read the answer straight from the model's probabilities: one pass over the text, nothing generated and nothing to parse. That makes them fast enough to check every step of an agent loop. Decido-1-Max and Decido-1-Think reason first, for the hardest calls.
The same answer, every time.
On CPU, Decido's outputs are bit-identical across calls and request history (at a fixed thread count), so any decision can be replayed exactly for an audit. It is also far less sensitive to the order in which answer options are listed.
Probabilities are calibrated, so a threshold can be set like an error budget: act automatically when the model is sure, and ask a person when it isn't.
A document and typed questions in. Probabilities out.
copy, paste, done ↓
# pip install decido-ai from decido import Decido ticket = "Hi, I was charged twice for my March invoice. Can you refund the duplicate?" m = Decido("decido-ai/decido-1-fast-1.7b") # downloads once, then runs locally r = m.decide(ticket, { "refund": "Is the customer asking for a refund?", "team": {"q": "Which team?", "options": ["billing", "technical", "sales"]}, "urgency": {"q": "How urgent?", "scale": ["low", "medium", "high"]}, }) # r["refund"] -> {"yes": 0.98, "no": 0.02} r["team"] -> {"billing": 0.95, "technical": 0.02, "sales": 0.03} if r["refund"]["yes"] >= 0.95: issue_refund() # sure: act (your own functions) else: send_to_human() # unsure: ask a person
decido models # the six models, scores and latency decido run decido-ai/decido-1-fast-1.7b "Hi, I was charged twice for my March invoice." \ -q refund="Is the customer asking for a refund?" \ -q team="Which team?|billing,technical,sales" \ -q urgency="How urgent?|scale:low,medium,high"
decido serve decido-ai/decido-1-fast-1.7b --port 8080 # local API, Jev 1.13 wire format; keeps running # in another terminal: curl localhost:8080/v1/systemone -H "content-type: application/json" -d '{ "state": "Hi, I was charged twice for my March invoice. Can you refund the duplicate?", "questions": { "refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"}, "team": {"type": "choice", "instructions": "Which team?", "criteria": {"billing": "", "technical": "", "sales": ""}} } }' # -> {"answers": {"refund": {"type": "noul", "noul": 0.98}, # "team": {"type": "choice", "choice": "billing", "confidence": 0.92, # "probabilities": {"billing": 0.95, "technical": 0.02, "sales": 0.03}}}, ...} # Existing Jev SDK clients work unchanged: TypeSafeClient(base_url="http://127.0.0.1:8080", api_key="local") # (the SDK zstd-compresses large requests; for decido serve to accept them: pip install "decido-ai[server]")
pip install on Mac (Apple Silicon), Linux (x86_64, aarch64) and Windows 10+ (x86_64): nothing to compile. Quickstart → On a GPU: pip install "decido-ai[gpu]" and device="cuda" (Decido-1-Max: CPU and Mac GPU for now; GPU weights soon).Six open models, one interface.
Decido-1-Ultra
- Benchmark mean
- 62.2
- 5 questions, 1k tokens
- 0.41 s CPU · 0.13 s Mac GPU
- Hardware
- laptop · ~4 GB RAM
- Deterministic
- bit-identical
Decido-1-Fast
- Benchmark mean
- 73.0
- 5 questions, 1k tokens
- 0.64 s CPU · 0.30 s Mac GPU
- Hardware
- laptop · ~8 GB RAM
- Deterministic
- bit-identical
Decido-1
- Benchmark mean
- 85.4
- 5 questions, 1k tokens
- 4.1 s CPU · 1.1 s Mac GPU
- Hardware
- laptop · ~30 GB RAM · GPU
- Deterministic
- bit-identical
Decido-1-Think
- Benchmark mean
- 88.7
- Per question
- ~19 s CPU · ~10 s Mac GPU
- Hardware
- laptop · ~30 GB RAM · GPU
- Deterministic
- 0% on CPU and Mac GPU
Decido-1-Max
- Benchmark mean
- 89.2
- Per question
- ~28 s CPU · ~11 s Mac GPU
- Hardware
- laptop · ~30 GB RAM (GPU path coming)
- Calibration
- Brier ≤ Jev on 5/5
Decido-1-XL
- Benchmark mean
- 87.3
- Hardware
- 4× A100 80 GB
- Install
- decido-ai[xl]
How each model was evaluated: see the technical report.
How it performs, and where it falls short.
yes, we list the weak spots too
Decido 1: calibrated decision models for AI agents
Technical report: benchmarks, reliability and cost.
One call, three typed questions
Real output from Decido-1-Fast on a laptop CPU: an injection check, the customer's intent and urgency, with a probability for each.
What the numbers do and don't show
Published-vs-measured comparisons, speed of the reasoning models, and calibration outside benchmarks.
Own your agent's decisions.
We're working with a small number of design partners building AI agents. We fine-tune Decido on your own decisions and you run it on your own servers.