Deqio

Decision in, probabilities out.

What is Deqio

Deqio is a local server for decision models — small language models trained to answer a typed question about a piece of text with a constrained, probabilistic answer instead of free-form prose. You send a state (text or JSON) and a question; you get back a yes/no, exactly one option, an ordered score, a set of labels or a cost-aware action — always with the probability of every allowed answer.

A generative model might answer:

"I think the refund should probably be approved."

A decision model answers:

{"choice": "approve", "probabilities": {"approve": 0.91, "review": 0.07, "reject": 0.02}}

That makes Deqio useful wherever the allowed answers are known in advance and a large LLM would be slow, expensive or unpredictable: routing and triage, policy checks and guardrails, agent checkpoints (next step, stop or continue, ask a human), scoring and classification. Deqio keeps the differences between model families and runtimes behind one HTTP API, records provenance for every answer, and ships a browser UI, a live request view, a benchmark runner and a Python client. Everything runs on your machine; the two remote decision APIs are optional catalog entries.

Install with uv

Install uv if you do not have it, then Deqio:

curl -LsSf https://astral.sh/uv/install.sh | sh
uv tool install deqio

Pick a workspace directory, install a model and start the server:

mkdir -p deqio-work && cd deqio-work
deqio models setup          # backend → model → token budget → download, install and probe
deqio serve                 # http://127.0.0.1:8787

Then open the dashboard at http://127.0.0.1:8787/ui, the live request view at /ui/watch and the API documentation at /docs. Deqio binds to 127.0.0.1 and keeps its state in the workspace (.deqio/, .model-runtimes/, models/, config.json).

Supported platforms: macOS on Apple Silicon (MLX and GGUF backends) and Linux (CUDA with an NVIDIA GPU, GGUF). Python 3.10 and 3.12 are tested in CI. Windows is not supported.

Models

Every model profile in the Deqio 0.6.5 catalog (31 models). A check mark means Deqio has a profile for that model/backend combination; availability on your machine still depends on the operating system, accelerator and memory. deqio models list prints the catalog for your host, deqio models installed what is ready to use.

Model	Model ID	MLX	CUDA	GGUF	Remote
SemIf / Qwen3.5 4B	`semif-qwen3.5-4b`	✓ 8-bit	✓	—	—
Decision 3.0 d3 26B	`d3`	—	✓	✓ Q8_0	—
Decision 3.0 d3-flash 8B	`d3-flash`	—	✓	✓ Q8_0	—
Decision 3.0 d3-mini 4.5B	`d3-mini`	—	✓	✓ Q8_0	—
Decision 3.0 d3-nano 2.2B	`d3-nano`	—	✓	✓ Q8_0	—
Decision 3.0 d3-lite 0.85B	`d3-lite`	—	✓	✓ Q8_0	—
Kev 0.8B	`kev-0.8b`	✓	✓	✓ Q8_0	—
Kev 4B	`kev-4b`	✓	✓	✓ Q8_0	—
Kev 9B	`kev-9b`	✓	✓	✓ Q8_0	—
Kev 27B	`kev-27b`	—	✓	—	—
JevK5 4B	`jevk5-4b`	—	✓	✓ Q8_0	—
JevK5 9B	`jevk5-9b`	—	✓	✓ Q8_0	—
Open-Jev 2B	`open-jev-2b`	—	✓	—	—
Open-Jev 9B	`open-jev-9b`	—	✓	—	—
Open-Jev 27B v1.1	`open-jev-27b-v1.1`	—	✓	—	—
CLM 8B	`clm-8b`	—	✓	—	—
Basal 2.0 Mini — 1.5B	`basal-2.0-mini`	✓ 8-bit	—	✓ Q8_0	—
Basal 2.0 — 4.5B	`basal-2.0-main`	✓ 8-bit	—	✓ Q8_0	—
Basal 2.0 Max — 11B	`basal-2.0-max`	✓ 8-bit	—	✓ Q8_0	—
Clef Flash 9B	`clef-flash`	✓ 8-bit	—	✓ Q8_0	—
Clef 27B	`clef`	✓ 8-bit	—	✓ Q8_0	—
Decider 0.8B	`decider-0.8b`	—	✓	—	—
Decider 2B	`decider-2b`	—	✓	✓ Q8_0	—
Decider 4B	`decider-4b`	—	✓	✓ Q8_0	—
Laya English 421M	`laya-english`	✓	✓	✓ Q8_0	—
Laya Multilingual 322M	`laya-multilingual`	✓	✓	—	—
Laya Typed Decisions 421M	`laya-typed-decisions`	✓	✓	—	—
Von	`von`	—	✓	—	—
Bespoke Nimble 9B	`nimble-9b`	✓	✓	—	—
TypeSafe Jev (remote)	`typesafe-jev`	—	—	—	✓
OpenAI Decisions API (remote)	`openai-decisions`	—	—	—	✓

✓ 8-bit is the author's MLX 8-bit checkpoint, ✓ Q8_0 the official GGUF Q8_0 file through the pinned llama.cpp or the engine's own GGUF sidecar, ✓ under Remote a remote decision API (no weights; the API key lives in the OS keychain).
Every engine runs in its own isolated environment with every runtime package and model revision pinned; every profile is labelled official or community.
Capabilities differ by model: Choice, Noul and Score are native everywhere; Multi is native on Basal 2.0 and emulated on every other profile (multi_emulated — one yes/no branch per label in one engine request); Act, Facts and option keys are Basal 2.0 features. Unsupported features are refused explicitly, never silently ignored.
Backends and model families in detail, remote API setup, token budgets and the deqio models commands: docs/MODELS.md.

Benchmarks and results

Deqio ships two editable benchmark suites — ENG Bench (deqio-basic-220: 55 Noul, 55 Choice, 55 Shared and 55 Multi cases) and PL Bench (deqio-pl-120: 30 of each, Polish-native, not a translation) — covering customer service, retail, security triage, AI-agent checkpoints, office administration, politics and world news. Run them against your installed models with:

deqio benchmark                     # choose a suite and the profiles interactively
deqio benchmark --all               # every installed profile; results under .deqio/benchmarks/<timestamp>
deqio benchmark compare             # two completed runs side by side

How to read the tables: Accuracy is the share of whole cases answered exactly right; Decision accuracy scores every decision — each Shared decision and each Multi label counts once. A Multi case passes only when all of its 6–8 labels are right, so its case accuracy is roughly the label accuracy to the 7th power; compare models on Multi label accuracy. Latencies are per request on one Apple Silicon laptop with one model resident at a time; smaller groups ran separately from larger ones. The two runtime errors in the above-3B ENG run are two SemIf cases refused by its upstream shared-state prefix check, reported as errors, not worked around. Everything else about the suites, the scoring and the tables is in docs/BENCHMARKS.md; the complete reports with the previous snapshots are benchmarks/summary_eng.md and benchmarks/summary_pl.md.

0.6.4 snapshot

Three groups measured on PL Bench and ENG Bench with Deqio 0.6.4 on 2026-10-18 (0.6.5 is the same code, published after the interrupted 0.6.4 upload): models up to 1B (d3-lite GGUF, Kev 0.8B, Laya English), models from 1B to 3B (d3-nano GGUF, Decider 2B, Basal 2.0 Mini) and models above 3B (d3-mini GGUF, Decider 4B, Basal 2.0 on GGUF and on MLX, JevK5 4B, Kev 4B, SemIf). The tables are rendered by deqio benchmark tables from the committed runs/<group>-<lang>/summary.json files; the remote decision APIs were last measured on the v1 PL suite (see the PL report).

Models up to 1B — PL Bench (0.6.4 snapshot)

Run

20261018T080338Z · suite deqio-pl-120 — 120 cases / 344 scored decisions · Deqio 0.6.4 · 2026-10-18
runtime errors across the listed profiles: 0

Quality

Accuracy is whole-case accuracy; Decision accuracy scores every decision/assertion (a Shared case holds several; a Multi case scores every label as present or absent).

Model	Backend	Accuracy	Decision accuracy	Noul	Choice	Shared	Multi	Shared decision accuracy	Multi label accuracy
Decision 3.0 d3-lite 0.85B	GGUF Q8_0	68.3%	82.8%	83.3%	80.0%	76.7%	33.3%	91.6%	79.6%
Kev 0.8B	MLX	52.5%	78.5%	66.7%	66.7%	56.7%	20.0%	83.1%	80.1%
Laya English 421M	MLX	26.7%	56.1%	46.7%	43.3%	13.3%	3.3%	57.8%	58.7%

Performance

Median and P95 are per-request latencies; throughput is scored decisions per second over the whole run; cost is the estimated USD per 1 000 scored decisions from the API's reported input tokens and its published input price (remote profiles only; an estimate, not a bill).

Model	Backend	Median	P95	Throughput	Cost / 1k decisions
Decision 3.0 d3-lite 0.85B	GGUF Q8_0	135.9 ms	592.7 ms	11.60/s	—
Kev 0.8B	MLX	86.2 ms	171.0 ms	29.94/s	—
Laya English 421M	MLX	67.8 ms	285.6 ms	23.84/s	—

Identity

Model	Backend	Engine	Runtime	Precision	Source	Model / endpoint	Revision
Decision 3.0 d3-lite 0.85B	GGUF Q8_0	decision3	llama.cpp-systemone	q8_0	official	`ggml-org/vllm-d3-lite-GGUF`	a1f9fa12
Kev 0.8B	MLX	kev	—	—	official	`jaredpalmer/kev-0.8b@bf75a6a8848ea6960ff2ed108d9ed44c2941174f`	bf75a6a8
Laya English 421M	MLX	laya	—	—	community	`aac6fef/laya-mlx`	04767856

Models up to 1B — ENG Bench (0.6.4 snapshot)

Run

20261018T080133Z · suite deqio-basic-220 — 220 cases / 652 scored decisions · Deqio 0.6.4 · 2026-10-18
runtime errors across the listed profiles: 0

Quality

Accuracy is whole-case accuracy; Decision accuracy scores every decision/assertion (a Shared case holds several; a Multi case scores every label as present or absent).

Model	Backend	Accuracy	Decision accuracy	Noul	Choice	Shared	Multi	Shared decision accuracy	Multi label accuracy
Decision 3.0 d3-lite 0.85B	GGUF Q8_0	75.0%	86.2%	92.7%	94.5%	70.9%	41.8%	88.4%	83.1%
Kev 0.8B	MLX	61.8%	80.8%	81.8%	87.3%	58.2%	20.0%	82.9%	78.8%
Laya English 421M	MLX	45.9%	73.5%	63.6%	65.5%	27.3%	27.3%	65.9%	79.4%

Performance

Median and P95 are per-request latencies; throughput is scored decisions per second over the whole run; cost is the estimated USD per 1 000 scored decisions from the API's reported input tokens and its published input price (remote profiles only; an estimate, not a bill).

Model	Backend	Median	P95	Throughput	Cost / 1k decisions
Decision 3.0 d3-lite 0.85B	GGUF Q8_0	137.1 ms	546.6 ms	13.55/s	—
Kev 0.8B	MLX	75.0 ms	144.4 ms	36.45/s	—
Laya English 421M	MLX	47.1 ms	166.7 ms	41.01/s	—

Identity

Model	Backend	Engine	Runtime	Precision	Source	Model / endpoint	Revision
Decision 3.0 d3-lite 0.85B	GGUF Q8_0	decision3	llama.cpp-systemone	q8_0	official	`ggml-org/vllm-d3-lite-GGUF`	a1f9fa12
Kev 0.8B	MLX	kev	—	—	official	`jaredpalmer/kev-0.8b@bf75a6a8848ea6960ff2ed108d9ed44c2941174f`	bf75a6a8
Laya English 421M	MLX	laya	—	—	community	`aac6fef/laya-mlx`	04767856

Models from 1B to 3B — PL Bench (0.6.4 snapshot)

Run

20261018T082457Z · suite deqio-pl-120 — 120 cases / 344 scored decisions · Deqio 0.6.4 · 2026-10-18
runtime errors across the listed profiles: 0

Quality

Accuracy is whole-case accuracy; Decision accuracy scores every decision/assertion (a Shared case holds several; a Multi case scores every label as present or absent).

Model	Backend	Accuracy	Decision accuracy	Noul	Choice	Shared	Multi	Shared decision accuracy	Multi label accuracy
Decision 3.0 d3-nano 2.2B	GGUF Q8_0	78.3%	89.0%	93.3%	93.3%	86.7%	40.0%	95.2%	85.1%
Basal 2.0 Mini — 1.5B	MLX 8-bit	75.8%	87.8%	83.3%	90.0%	86.7%	43.3%	95.2%	85.1%
Decider 2B	GGUF Q8_0	74.2%	87.5%	90.0%	90.0%	83.3%	33.3%	92.8%	84.6%

Performance

Median and P95 are per-request latencies; throughput is scored decisions per second over the whole run; cost is the estimated USD per 1 000 scored decisions from the API's reported input tokens and its published input price (remote profiles only; an estimate, not a bill).

Model	Backend	Median	P95	Throughput	Cost / 1k decisions
Decision 3.0 d3-nano 2.2B	GGUF Q8_0	320.7 ms	1,420.1 ms	4.86/s	—
Basal 2.0 Mini — 1.5B	MLX 8-bit	232.3 ms	496.2 ms	10.24/s	—
Decider 2B	GGUF Q8_0	225.9 ms	878.9 ms	7.83/s	—

Identity

Model	Backend	Engine	Runtime	Precision	Source	Model / endpoint	Revision
Decision 3.0 d3-nano 2.2B	GGUF Q8_0	decision3	llama.cpp-systemone	q8_0	official	`ggml-org/vllm-d3-nano-GGUF`	67baf946
Basal 2.0 Mini — 1.5B	MLX 8-bit	basal	basal-2.0-mlx 2.0.0	mlx-8bit	official	`Remek/basal-2.0-mini-MLX-8bit`	f26d1e2d
Decider 2B	GGUF Q8_0	decider	decider-gguf-native	q8_0	official	`Mapika/decider-2b-GGUF`	ff2e5e68

Models from 1B to 3B — ENG Bench (0.6.4 snapshot)

Run

20261018T081947Z · suite deqio-basic-220 — 220 cases / 652 scored decisions · Deqio 0.6.4 · 2026-10-18
runtime errors across the listed profiles: 0

Quality

Accuracy is whole-case accuracy; Decision accuracy scores every decision/assertion (a Shared case holds several; a Multi case scores every label as present or absent).

Model	Backend	Accuracy	Decision accuracy	Noul	Choice	Shared	Multi	Shared decision accuracy	Multi label accuracy
Decision 3.0 d3-nano 2.2B	GGUF Q8_0	80.0%	90.5%	96.4%	90.9%	81.8%	50.9%	92.7%	88.6%
Decider 2B	GGUF Q8_0	79.5%	90.5%	90.9%	92.7%	83.6%	50.9%	93.9%	88.6%
Basal 2.0 Mini — 1.5B	MLX 8-bit	73.6%	87.9%	78.2%	92.7%	69.1%	54.5%	85.4%	89.7%

Performance

Median and P95 are per-request latencies; throughput is scored decisions per second over the whole run; cost is the estimated USD per 1 000 scored decisions from the API's reported input tokens and its published input price (remote profiles only; an estimate, not a bill).

Model	Backend	Median	P95	Throughput	Cost / 1k decisions
Decision 3.0 d3-nano 2.2B	GGUF Q8_0	333.5 ms	1,394.1 ms	5.58/s	—
Decider 2B	GGUF Q8_0	213.8 ms	796.7 ms	9.79/s	—
Basal 2.0 Mini — 1.5B	MLX 8-bit	314.2 ms	629.8 ms	8.41/s	—

Identity

Model	Backend	Engine	Runtime	Precision	Source	Model / endpoint	Revision
Decision 3.0 d3-nano 2.2B	GGUF Q8_0	decision3	llama.cpp-systemone	q8_0	official	`ggml-org/vllm-d3-nano-GGUF`	67baf946
Decider 2B	GGUF Q8_0	decider	decider-gguf-native	q8_0	official	`Mapika/decider-2b-GGUF`	ff2e5e68
Basal 2.0 Mini — 1.5B	MLX 8-bit	basal	basal-2.0-mlx 2.0.0	mlx-8bit	official	`Remek/basal-2.0-mini-MLX-8bit`	f26d1e2d

Models above 3B — PL Bench (0.6.4 snapshot)

Run

20261018T093333Z · suite deqio-pl-120 — 120 cases / 344 scored decisions · Deqio 0.6.4 · 2026-10-18
runtime errors across the listed profiles: 0

Quality

Accuracy is whole-case accuracy; Decision accuracy scores every decision/assertion (a Shared case holds several; a Multi case scores every label as present or absent).

Model	Backend	Accuracy	Decision accuracy	Noul	Choice	Shared	Multi	Shared decision accuracy	Multi label accuracy
Decider 4B	GGUF Q8_0	80.8%	92.4%	93.3%	93.3%	96.7%	40.0%	98.8%	89.6%
Decision 3.0 d3-mini 4.5B	GGUF Q8_0	82.5%	92.2%	96.7%	96.7%	93.3%	43.3%	97.6%	88.6%
Basal 2.0 — 4.5B	MLX 8-bit	80.0%	91.6%	90.0%	93.3%	83.3%	53.3%	92.8%	91.0%
JevK5 4B	GGUF Q8_0	82.5%	90.4%	96.7%	96.7%	96.7%	40.0%	98.8%	85.1%
SemIf / Qwen3.5 4B	MLX	80.8%	89.8%	96.7%	96.7%	90.0%	40.0%	96.4%	85.1%
Kev 4B	GGUF Q8_0	75.0%	82.8%	93.3%	96.7%	93.3%	16.7%	97.6%	73.1%

Performance

Median and P95 are per-request latencies; throughput is scored decisions per second over the whole run; cost is the estimated USD per 1 000 scored decisions from the API's reported input tokens and its published input price (remote profiles only; an estimate, not a bill).

Model	Backend	Median	P95	Throughput	Cost / 1k decisions
Decider 4B	GGUF Q8_0	560.5 ms	2,299.8 ms	3.12/s	—
Decision 3.0 d3-mini 4.5B	GGUF Q8_0	1,031.1 ms	4,694.0 ms	1.56/s	—
Basal 2.0 — 4.5B	MLX 8-bit	953.5 ms	1,964.7 ms	2.62/s	—
JevK5 4B	GGUF Q8_0	1,188.2 ms	5,235.8 ms	1.39/s	—
SemIf / Qwen3.5 4B	MLX	968.1 ms	3,005.5 ms	2.03/s	—
Kev 4B	GGUF Q8_0	713.4 ms	1,672.9 ms	3.54/s	—

Identity

Model	Backend	Engine	Runtime	Precision	Source	Model / endpoint	Revision
Decider 4B	GGUF Q8_0	decider	decider-gguf-native	q8_0	official	`Mapika/decider-4b-GGUF`	b79f09d9
Decision 3.0 d3-mini 4.5B	GGUF Q8_0	decision3	llama.cpp-systemone	q8_0	official	`ggml-org/vllm-d3-mini-GGUF`	c03ac9c3
Basal 2.0 — 4.5B	MLX 8-bit	basal	basal-2.0-mlx 2.0.0	mlx-8bit	official	`Remek/basal-2.0-4.5B-MLX-8bit`	77d3b723
JevK5 4B	GGUF Q8_0	jevk5	jevk5-gguf-native	q8_0	official	`alibiserikbay/JevK5-GGUF`	ec67b0bf
SemIf / Qwen3.5 4B	MLX	semif	—	—	official	`Qwen/Qwen3.5-4B`	851bf6e8
Kev 4B	GGUF Q8_0	kev	llama.cpp-systemone	q8_0	official	`ggml-org/Kev-4B-GGUF`	d924f2e2

Models above 3B — ENG Bench (0.6.4 snapshot)

Run

20261018T085501Z · suite deqio-basic-220 — 220 cases / 652 scored decisions · Deqio 0.6.4 · 2026-10-18
runtime errors across the listed profiles: 2

Quality

Accuracy is whole-case accuracy; Decision accuracy scores every decision/assertion (a Shared case holds several; a Multi case scores every label as present or absent).

Model	Backend	Accuracy	Decision accuracy	Noul	Choice	Shared	Multi	Shared decision accuracy	Multi label accuracy
Decision 3.0 d3-mini 4.5B	GGUF Q8_0	89.1%	95.7%	96.4%	100.0%	96.4%	63.6%	98.8%	93.7%
Decider 4B	GGUF Q8_0	86.8%	94.6%	98.2%	98.2%	90.9%	60.0%	97.0%	92.6%
Basal 2.0 — 4.5B	GGUF Q8_0	85.5%	94.2%	89.1%	96.4%	89.1%	67.3%	96.3%	93.7%
Basal 2.0 — 4.5B	MLX 8-bit	85.5%	94.2%	89.1%	96.4%	90.9%	65.5%	97.0%	93.4%
JevK5 4B	GGUF Q8_0	85.5%	92.9%	96.4%	98.2%	96.4%	50.9%	98.8%	89.2%
Kev 4B	GGUF Q8_0	80.5%	91.7%	96.4%	100.0%	78.2%	47.3%	92.1%	89.7%
SemIf / Qwen3.5 4B	MLX	80.5%	88.2%	98.2%	96.4%	87.3%	40.0%	95.1%	82.5%

Performance

Median and P95 are per-request latencies; throughput is scored decisions per second over the whole run; cost is the estimated USD per 1 000 scored decisions from the API's reported input tokens and its published input price (remote profiles only; an estimate, not a bill).

Model	Backend	Median	P95	Throughput	Cost / 1k decisions
Decision 3.0 d3-mini 4.5B	GGUF Q8_0	1,096.4 ms	5,012.8 ms	1.61/s	—
Decider 4B	GGUF Q8_0	525.7 ms	2,068.8 ms	3.91/s	—
Basal 2.0 — 4.5B	GGUF Q8_0	1,243.1 ms	2,059.4 ms	2.43/s	—
Basal 2.0 — 4.5B	MLX 8-bit	1,248.9 ms	2,481.6 ms	2.13/s	—
JevK5 4B	GGUF Q8_0	1,491.3 ms	5,341.8 ms	1.40/s	—
Kev 4B	GGUF Q8_0	663.7 ms	1,576.8 ms	3.90/s	—
SemIf / Qwen3.5 4B	MLX	1,092.7 ms	3,165.8 ms	2.07/s	—

Identity

Model	Backend	Engine	Runtime	Precision	Source	Model / endpoint	Revision
Decision 3.0 d3-mini 4.5B	GGUF Q8_0	decision3	llama.cpp-systemone	q8_0	official	`ggml-org/vllm-d3-mini-GGUF`	c03ac9c3
Decider 4B	GGUF Q8_0	decider	decider-gguf-native	q8_0	official	`Mapika/decider-4b-GGUF`	b79f09d9
Basal 2.0 — 4.5B	GGUF Q8_0	basal	basal-2.0-gguf 2.0.0	q8_0	official	`Remek/basal-2.0-4.5B`	e5ba9149, 237294ca
Basal 2.0 — 4.5B	MLX 8-bit	basal	basal-2.0-mlx 2.0.0	mlx-8bit	official	`Remek/basal-2.0-4.5B-MLX-8bit`	77d3b723
JevK5 4B	GGUF Q8_0	jevk5	jevk5-gguf-native	q8_0	official	`alibiserikbay/JevK5-GGUF`	ec67b0bf
Kev 4B	GGUF Q8_0	kev	llama.cpp-systemone	q8_0	official	`ggml-org/Kev-4B-GGUF`	d924f2e2
SemIf / Qwen3.5 4B	MLX	semif	—	—	official	`Qwen/Qwen3.5-4B`	851bf6e8

Install from source

git clone https://github.com/ILuce/deqio.git && cd deqio
uv sync --frozen --extra dev        # Deqio core + the deqio-client SDK (workspace member)
uv run ruff check src tests clients
uv run pytest
uv run deqio models setup
uv run deqio serve

CI (.github/workflows/ci.yml) runs the same lint, byte-compile and test steps with Python 3.10 and 3.12. Release notes: CHANGELOG.md; release procedure: docs/RELEASE_CHECKLIST.md; security reporting: SECURITY.md; the engineering rules of the project: AGENTS.md.

Use the API

The server listens on http://127.0.0.1:8787. Every route takes and returns JSON; the full reference with an example per route is docs/API.md, and /docs renders the OpenAPI document of the running server.

Route	Answers	Use it for
`POST /v1/noul`	yes / no with probabilities	true/false checks, guardrails
`POST /v1/choice`	exactly one option	routing, classification, triage (`/v1/decision` is an alias)
`POST /v1/shared`	several decisions over one state, one engine call	checklists, multi-view validation
`POST /v1/score`	one level of an ordered scale	severity, risk, urgency
`POST /v1/multi`	a set of labels, one probability each	tagging, multi-label classification
`POST /v1/act`	the action with the lowest expected cost	decisions where mistakes have different prices
`POST /v1/soam`	one state, many typed questions, one native request	agent checkpoints, release validation
`POST /v1/systemone`	the same, in the raw System One contract	low-level integrations, adapters

noul, choice and shared are the portable routes: the same request and response on every model. score, multi, act, soam and systemone are the native typed routes: they forward the model's own System One contract and are available when the active profile supports them — deqio models installed shows the capability matrix, and a feature the model lacks is refused with 422, never silently ignored.

A yes/no question:

curl -s http://127.0.0.1:8787/v1/noul -H 'content-type: application/json' -d '{
  "state": "The patch changed authentication code. Unit tests passed but integration tests have not run.",
  "question": "Should integration tests run before release?"
}'

Exactly one option:

curl -s http://127.0.0.1:8787/v1/choice -H 'content-type: application/json' -d '{
  "state": "A customer was charged twice for one subscription renewal.",
  "question": "Which team should handle this request?",
  "options": [
    {"id": "billing", "description": "Payments, refunds and invoices."},
    {"id": "support", "description": "Product and account support."},
    {"id": "sales", "description": "New purchases and upgrades."}
  ]
}'

One state, several typed questions in one native request:

curl -s http://127.0.0.1:8787/v1/soam -H 'content-type: application/json' -d '{
  "state": "The laptop arrived with a cracked screen and the customer requests a refund.",
  "questions": {
    "route":    {"type": "choice", "instructions": "Which workflow handles this?",
                 "criteria": {"refund": "Refund workflow.", "repair": "Repair workflow.", "review": "Human review."}},
    "damaged":  {"type": "noul", "instructions": "Is the item damaged?"},
    "severity": {"type": "score", "instructions": "Rate severity.", "criteria": ["low", "medium", "high"]}
  }
}'

Every answer carries probabilities and a provenance block (engine, model, backend, pinned revisions, and the provider, request id and cost estimate for a remote profile), so a decision can always be tied to the exact model that made it. Operational routes: /health, /v1/models, /v1/models/installed, /v1/models/activate, /v1/recent, /v1/watch.

Use the Python client

deqio-client is a dependency-free Python SDK, installed together with Deqio (or alone with pip install deqio-client). One method per route; every method returns the server's JSON unchanged and raises DeqioError with the status and detail when the server rejects the request:

from deqio_client import DeqioClient, DeqioError

client = DeqioClient("http://127.0.0.1:8787")

degraded = client.noul("Checkout works after retry, but payment latency is elevated.", "Is checkout degraded?")
report = client.choice("The API returns HTTP 500 after the latest deployment.", "Classify the report.",
                       {"bug": "An existing function no longer works", "feature": "A request for new functionality"})
severity = client.score("Payment latency is elevated for some customers.",
                        {"instructions": "Rate the operational severity.", "criteria": ["low", "medium", "high", "critical"]})
views = client.soam("One confirmed state.", {"ok": {"type": "noul", "instructions": "Is it fine?"},
                                            "kind": {"type": "choice", "criteria": {"a": "A", "b": "B"}}})

print(report["decision"], report["probabilities"])
print(severity["answers"]["decision"], degraded["provenance"]["engine"])

client.shared, client.multi, client.act and client.systemone cover the remaining routes; contract=True negotiates the fail-closed input-completeness contract on the portable routes. The same calls are available from the shell through the SDK:

deqio decide noul   --state "Payment latency is elevated." --question "Is checkout degraded?"
deqio decide choice --state "API returns 500 after deploy." --question "Classify." --option bug="Broken function" --option feature="New request"
deqio decide multi  --state "Cracked screen, damaged box, refund asked." --criterion damage="Product damage" --criterion refund="Refund requested"
deqio decide soam   --state-file state.txt --questions-json @questions.json --json

Summary

Deqio turns a decision model into a local service: one API for yes/no, single-choice, scored, multi-label and cost-aware decisions, 31 models from 322M to 27B parameters on MLX, GGUF and CUDA plus two remote APIs, every answer with probabilities and provenance, a browser dashboard and live request view, and benchmark suites in English and Polish whose current results you have just read. Install it with uv tool install deqio, pick a model with deqio models setup, and start with /v1/choice or the Python client. Operational details live in docs/API.md, docs/MODELS.md and docs/BENCHMARKS.md.

License

MIT. See LICENSE.

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
benchmarks		benchmarks
clients/python		clients/python
docs		docs
logs		logs
models		models
runs		runs
src/deqio		src/deqio
tests		tests
.gitignore		.gitignore
.python-version		.python-version
AGENTS.md		AGENTS.md
CHANGELOG.md		CHANGELOG.md
LICENSE		LICENSE
README.md		README.md
SECURITY.md		SECURITY.md
models.json		models.json
pyproject.toml		pyproject.toml
uv.lock		uv.lock

Latest commit

History

Folders and files

Repository files navigation

Deqio

What is Deqio

Install with uv

Models

Benchmarks and results

0.6.4 snapshot

Models up to 1B — PL Bench (0.6.4 snapshot)

Run

Quality

Performance

Identity

Models up to 1B — ENG Bench (0.6.4 snapshot)

Run

Quality

Performance

Identity

Models from 1B to 3B — PL Bench (0.6.4 snapshot)

Run

Quality

Performance

Identity

Models from 1B to 3B — ENG Bench (0.6.4 snapshot)

Run

Quality

Performance

Identity

Models above 3B — PL Bench (0.6.4 snapshot)

Run

Quality

Performance

Identity

Models above 3B — ENG Bench (0.6.4 snapshot)

Run

Quality

Performance

Identity

Install from source

Use the API

Use the Python client

Summary

License

About

Topics

Resources

Security policy

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages