Typed decisions.
Immediately.
Statim runs decision models as a native C++ engine. Ask any text a choice, a score or a yes/no question, and get calibrated answers from one forward pass on your own CPU or GPU.
- 5.5 MB static binary
- No Python at runtime
- Apache-2.0 engine
department choice
Which department should handle this?
- billing0.978
- sales0.016
- technical0.006
Answer billing
urgency score
How urgent is this request?
Probabilities: not urgent 0.24, soon 0.24, critical 0.52.
Score 1.28 of 2
refund yes / no
Does the user explicitly request a refund?
- yes0.915
- no0.085
Answer yes
All three questions in one forward pass: 190 tokens, about 0.3 s on a Ryzen 7 5800X CPU for the 395M-parameter model. Real output, not a mock-up.
Watch it decide
One support ticket, three typed answers, one forward pass: that is the film. The story shows where typed decisions replace rule trees.
Two models, one engine
The engine runs any Laya checkpoint. The Statim Decide models are fine-tuned on licence-audited data and released only after the no-harm gate below. Both are on Hugging Face as f32 and q8_0 files.
-
Statim Decide EN Large
ModernBERT-large, 395M parameters, English. Release 0.5.0.
- typed-decisions
- 0.768
- Banking77
- 0.928
- MASSIVE, English
- 0.867
-
Statim Decide Multilingual Base
mmBERT-base, evaluated in 12 languages and 14 decision categories. Release 0.10.0, noncommercial licence.
- typed-decisions
- 0.775
- Banking77
- 0.918
- Decision categories
- 0.826
Try both without installing anything in the live demo, or route real support tickets in ten minutes with the ticket triage example.
Measured against the alternatives
The same model and tokens against ONNX Runtime and the Laya PyTorch reference on one CPU, and S1Bench against Lev and Jev with Lev's own harness. Where Statim is behind, the chart says so.
Accuracy you can check
Every number here comes from a script in the repository. The best Statim Decide model on each benchmark next to published systems under the same protocol, and the held-out suites every release has to pass.
typed-decisions
0.768 Test split, 2,000 decisions. In-domain: its train split is replay data.- Statim Decide EN Large0.768
- meraGPT0.768
- laya-typed-decisions0.766
- Jev0.727
AG News, zero-shot
0.939 Never trained on. First 2,000 test rows, four topics.- Laya0.950
- Statim Decide EN Large0.939
- GPT-3 (CARP)0.926
- Jev0.881
Banking77
0.928 Trained on the train split, all 77 intents in one question.- MPNet, supervised classifier0.941
- Statim Decide EN Large0.928
MASSIVE, 12 languages
0.816 Trained on 2,000 rows per language; the paper's XLM-R used the full train set.- XLM-R base, MASSIVE paper0.857
- Statim Decide Multilingual0.816
Published results from the respective papers and model cards, collected with sources in docs/ROADMAP.md. Statim's numbers: tools/finetune/gate.py. Against a zero-shot Qwen3-8B and a zero-shot NLI classifier on the same 14 decision categories, Statim scores 0.826 macro accuracy to their 0.706 and 0.488; see docs/BASELINES.md.
No release without the no-harm gate
A new model replaces the old one only if validation holds (at most one point lower), no held-out suite drops significantly in a paired test on the same items after correcting for the number of suites, and it is measurably better somewhere. Here is Statim Decide Multilingual 0.10.0 against 0.7.0, the release it replaces, one square per suite, change in accuracy points. The English model passed the same gate: 10 significant gains, 0 regressions.
Trained tasks 14 suites
- Banking77: 0.914 to 0.918, within noise+0
- typed-decisions: 0.763 to 0.775, within noise+1
- MASSIVE intents ar: 0.707 to 0.740, within noise+3
- MASSIVE intents de: 0.787 to 0.827, within noise+4
- MASSIVE intents en: 0.827 to 0.847, within noise+2
- MASSIVE intents es: 0.813 to 0.833, within noise+2
- MASSIVE intents fr: 0.820 to 0.807, within noise−1
- MASSIVE intents hi: 0.787 to 0.780, within noise−1
- MASSIVE intents it: 0.820 to 0.813, within noise−1
- MASSIVE intents ja: 0.827 to 0.833, within noise+1
- MASSIVE intents pl: 0.800 to 0.833, within noise+3
- MASSIVE intents ru: 0.833 to 0.847, within noise+1
- MASSIVE intents tr: 0.787 to 0.813, within noise+3
- MASSIVE intents zh-CN: 0.787 to 0.820, within noise+3
Decision categories 37 suites
- Complaint en: 0.767 to 0.820, within noise+5
- Emotion de: 0.453 to 0.487, within noise+3
- Emotion en: 0.573 to 0.640, within noise+7
- Emotion es: 0.573 to 0.620, within noise+5
- Emotion fr: 0.660 to 0.827, significant gain+17
- Emotion hi: 0.713 to 0.860, significant gain+15
- Emotion pt: 0.507 to 0.560, within noise+5
- Emotion ru: 0.693 to 0.740, within noise+5
- Emotion zh: 0.540 to 0.553, within noise+1
- Fact-check en: 0.313 to 0.467, significant gain+15
- Formality ja: 0.547 to 0.907, significant gain+36
- Formality tr: 1.000 to 1.000, within noise+0
- Intent en: 0.820 to 0.847, within noise+3
- Intent nl: 0.440 to 0.533, within noise+9
- Intent tr: 1.000 to 1.000, within noise+0
- NLI en: 0.807 to 0.800, within noise−1
- NLI ja: 0.633 to 0.660, within noise+3
- NLI tr: 0.800 to 0.847, within noise+5
- PII ar: 0.847 to 0.933, within noise+9
- PII de: 0.840 to 0.887, within noise+5
- PII en: 0.893 to 0.933, within noise+4
- PII es: 0.853 to 0.887, within noise+3
- PII fr: 0.893 to 0.913, within noise+2
- PII it: 0.840 to 0.900, within noise+6
- PII ja: 0.847 to 0.927, within noise+8
- PII nl: 0.760 to 0.853, within noise+9
- PII ru: 0.907 to 0.967, within noise+6
- PII sv: 0.833 to 0.873, within noise+4
- PII zh: 0.900 to 0.927, within noise+3
- Reading en: 0.927 to 0.933, within noise+1
- Safety en: 0.727 to 0.880, significant gain+15
- Sentiment en: 0.800 to 0.853, within noise+5
- Sentiment zh: 0.800 to 0.900, within noise+10
- Similarity pt: 0.833 to 0.880, within noise+5
- Stance en: 0.893 to 0.980, within noise+9
- Topic en: 0.607 to 0.647, within noise+4
- Urgency en: 0.893 to 0.987, significant gain+9
Never trained on 40 suites
- AG News: 0.929 to 0.921, within noise−1
- DAIR Emotion: 0.504 to 0.530, within noise+3
- Belebele reading ar: 0.293 to 0.260, within noise−3
- Belebele reading de: 0.213 to 0.320, within noise+11
- Belebele reading en: 0.247 to 0.313, within noise+7
- Belebele reading hi: 0.280 to 0.227, within noise−5
- FarsTail NLI fa: 0.700 to 0.653, within noise−5
- GoEmotions en: 0.473 to 0.453, within noise−2
- HWU64 intents en: 0.787 to 0.853, within noise+7
- IndoNLI id: 0.667 to 0.700, within noise+3
- HateCheck ar: 0.573 to 0.667, within noise+9
- HateCheck de: 0.687 to 0.673, within noise−1
- HateCheck en: 0.613 to 0.647, within noise+3
- HateCheck es: 0.620 to 0.647, within noise+3
- HateCheck fr: 0.687 to 0.660, within noise−3
- HateCheck hi: 0.573 to 0.580, within noise+1
- HateCheck it: 0.660 to 0.620, within noise−4
- HateCheck nl: 0.653 to 0.647, within noise−1
- HateCheck pl: 0.633 to 0.653, within noise+2
- HateCheck pt: 0.647 to 0.673, within noise+3
- HateCheck zh: 0.647 to 0.640, within noise−1
- Sentiment ar: 0.560 to 0.633, within noise+7
- Sentiment de: 0.580 to 0.560, within noise−2
- Sentiment en: 0.693 to 0.720, within noise+3
- Sentiment es: 0.527 to 0.600, within noise+7
- Sentiment fr: 0.587 to 0.627, within noise+4
- Sentiment hi: 0.533 to 0.540, within noise+1
- Sentiment id: 0.747 to 0.780, within noise+3
- Sentiment it: 0.567 to 0.647, within noise+8
- Sentiment ja: 0.660 to 0.680, within noise+2
- Sentiment ms: 0.533 to 0.533, within noise+0
- Sentiment pt: 0.633 to 0.640, within noise+1
- Sentiment zh: 0.640 to 0.613, within noise−3
- SemRel similarity ar: 0.247 to 0.273, within noise+3
- SemRel similarity en: 0.200 to 0.280, within noise+8
- SemRel similarity hi: 0.227 to 0.293, within noise+7
- SIB-200 topics ar: 0.787 to 0.780, within noise−1
- SIB-200 topics de: 0.800 to 0.827, within noise+3
- SIB-200 topics en: 0.833 to 0.847, within noise+1
- SIB-200 topics hi: 0.707 to 0.733, within noise+3
6 significant gains85 within noise0 regressions
A re-audit on 2026-10-03 found eight sources in the mixture behind the released weights that are non-commercial, ShareAlike or unlicensed. Those weights, including 0.10.0, are therefore offered for noncommercial use only, and a model trained only on cleared sources is next. The training-source policy is now code and checked in CI. Every source, its licence and each finding are listed in DATA_LICENSES.md.
Same answers. A different machine.
Statim reimplements the Laya engine in C++20. Its answers match the Python reference on all 240 test sequences within 1e-4, and CI checks that on every change. What changes is everything around the forward pass.
-
Cold start to first answer
15× faster -
Server memory
6× less -
HTTP latency, p95
−74 % -
HTTP throughput
2.5× more
Laya is the Python reference server. Measured on a 2015-class laptop CPU with the multilingual checkpoint in f32, 30 states with 8 questions each; method and raw numbers in the README.
- Runtime
- Python 3.10+, PyTorch, transformersOne 5.5 MB binary and one .gguf file
- Peak memory, one request
- 2,667 MB585 MB in f32, 234 MB in q8_0
- Deploy
- pip or Docker with PyTorchStatic binary, distroless image or systemd unit
On a GPU, still exact
Statim on an RTX 3070 against the same machine's Ryzen 7 5800X, both in f32, with Vulkan or CUDA. The usual f16 shortcut stays off unless you pass --gpu-fast, so every answer stays identical. With --gpu-fast the English model serves 16.4 requests per second on Vulkan; q8_0 weights on CUDA reach 15.7.
- English model, per request12×1,683 ms on CPU, 137 ms on GPU
- Multilingual model, per request10×535 ms on CPU, 54 ms on GPU
- English over HTTP8.9×0.91 req/s on CPU, 8.1 req/s on GPU
- Multilingual over HTTP7.7×2.68 req/s on CPU, 20.5 req/s on GPU
One pass. Every option scored.
Other setups ask a model once per question, or once per option. Statim asks once.
Pack the request
The state and every question become one token sequence, with a [MASK] token for each option.
Run the model once
The encoder and the decision head run a single time, not once per question.
Read typed answers
Choice, score and yes/no answers come back calibrated, with confidence, token usage and routing.
- A model is one file
- The converter stores architecture, calibration and tokenizer in a single GGUF file. No sidecar vocabulary, no Python package.
- CPU first, GPU optional
- AVX2 and ARM NEON out of the box. Vulkan and CUDA backends are opt-in and exact by default.
- Two models, one answer
- Serve the English and the multilingual model together: route each request by language, or let both vote.
- Built for production
- Worker pool, admission control, bearer auth, Prometheus metrics, JSON logs, request IDs and graceful shutdown.
- Drop-in protocol
- Speaks the Jev/Laya
POST /v1/systemoneAPI. Existing clients switch by changing the base URL. - Hardened by default
- Limits on body size, JSON depth, tokens, attention memory and concurrency before any inference runs. Security regressions are tested in CI.
Speaks the Jev/Laya API
POST /v1/systemone returns {model, answers, usage, routing}, and errors keep FastAPI's {"detail": ...} shape. Requests can also select a loaded LoRA adapter by name, or "auto" to route by question category. Official Python and TypeScript clients use only the standard library and retry on 503.
- POST/v1/systemone
- One state, any number of questions
- POST/v1/systemone/batch
- Many states, packed into shared forward passes
- GET/v1/models
- Loaded models, their adapters and the device they run on
- GET/health, /ready
- Liveness and readiness probes
- GET/metrics
- Prometheus metrics
Reference: docs/API.md and the OpenAPI 3.1 spec.
curl
statim serve -m multilingual=models/laya-multilingual-f32.gguf --port 8080
curl -s localhost:8080/v1/systemone -d '{
"state": {"subject": "Duplicate charge on invoice #4411",
"body": "We were billed twice for March. Please refund the second charge."},
"questions": {
"department": {"type": "choice",
"instructions": "Which department should handle this?",
"criteria": {"billing": "invoices, refunds",
"technical": "bugs, outages",
"sales": "pricing, contracts"}},
"urgency": {"type": "score",
"instructions": "How urgent is this request?",
"criteria": ["not urgent", "soon", "critical"]},
"refund": {"type": "noul",
"instructions": "Does the user explicitly request a refund?"}}
}'Python
from statim import Client
client = Client("http://127.0.0.1:8080")
decision = client.decide(
{"subject": "Duplicate charge on invoice #4411",
"body": "We were billed twice for March. Please refund the second charge."},
{"department": {"type": "choice",
"instructions": "Which department should handle this?",
"criteria": {"billing": "invoices, refunds",
"technical": "bugs, outages"}},
"refund": {"type": "noul",
"instructions": "Does the user explicitly request a refund?"}},
)
print(decision.answers["department"].choice) # billing
print(decision.answers["refund"].yes) # TrueTypeScript
import { Client } from "@statim/client";
const client = new Client("http://127.0.0.1:8080");
const decision = await client.decide(
{ subject: "Duplicate charge on invoice #4411",
body: "We were billed twice for March. Please refund the second charge." },
{ department: { type: "choice",
instructions: "Which department should handle this?",
criteria: { billing: "invoices, refunds", technical: "bugs, outages" } },
refund: { type: "noul",
instructions: "Does the user explicitly request a refund?" } },
);
console.log(decision.answers.department.choice); // billingFrom clone to first answer
Build the static binary, fetch and convert the models, then serve both with consensus. The playground opens at http://127.0.0.1:8080/.
git clone --recursive https://github.com/BEKO2210/statim && cd statim
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release
cmake --build build
tools/fetch_models.sh multilingual english
./build/statim serve \
-m english=models/laya-english-f32.gguf \
-m multilingual=models/laya-multilingual-f32.gguf \
--consensus
- systemdHardened unit with memory and CPU quotas
- DockerDistroless image for CPU or Vulkan
- Probes/health, /ready and Prometheus /metrics
The engine is free. The weights have a commercial path.
The source is Apache-2.0, including for commercial products. Model weights published by Statim are free for noncommercial use, for small companies and for a 32-day trial, and licensed per organisation beyond that.
Source code
Free for any useApache License 2.0Engine, server and tools, in products, internal tools and SaaS.
Weights, noncommercial
FreePolyForm Noncommercial 1.0.0Research, study, hobby projects, education, charities and public institutions.
Weights, small companies and trials
Free, including commercial usePolyForm Small Business 1.0.0, Free Trial 1.0.0Companies below 100 people and 1 M USD revenue use them in production for free; any company may evaluate them for 32 days.
Weights, commercial
Per organisationStatim commercial licenceLarger companies in products, SaaS, paid APIs, internal tools and client work, with volume tiers for self-hosting.
See licensing