Total runs:
2.8K
Run Growth:
-3.7K
Growth Rate:
-129.92%
Sponsored by
APIMart
infosave / cmf-decision
Model's Last Updated: October 01 2026
text-classification
Introduction of cmf-decision
Model Details of cmf-decision
CMF Decision
Decisions in milliseconds. One portable CMF file.
Turn a customer message into a clear action: route a request, classify an intent, or choose a tool using your own trained skill. Cortiq selects a label or abstains when uncertain — without generating tokens locally . Connect an optional oracle for unfamiliar cases, then use its answers to grow your local skills.
Local decisions · Custom skills · Optional oracle
Try the live Space → — enter a request and inspect the decision, confidence and reconstruction errors. No installation or API key needed.
Quick start · Your own skill · Oracle · API · Benchmarks · На русском
Quick start
Use the hosted service
Prefer an API without local setup? allaigate Routing API is a ready-to-use routing service built on the CMF format. Get an API key on the site and follow the integration guide — no model download or server deployment required.
Run locally
Install the latest published CLI from crates.io (requires Rust 1.88 or newer):
cargo install cortiq-cli --locked
Run the same command again to update. Download the model and make your first decision:
curl -fL -o cortiq-decision.cmf \
https://huggingface.co/infosave/cmf-decision/resolve/main/cortiq-decision.cmf
cortiq decide cortiq-decision.cmf --skill banking77 -p "I still have not received my new card"
Result:
card_arrival
, answered locally. No cloud key required.
Included skills:
banking77
,
clinc150
,
massive
.
Choose your hardware
CPU works without configuration. The same installed CLI also includes GPU support (from Cortiq 0.8.0):
# Apple Silicon
CORTIQ_DECISION_DEVICE=metal cortiq decide cortiq-decision.cmf \
--skill banking77 -p "I still have not received my new card"
# Vulkan GPU (a hardware Vulkan driver is required)
CORTIQ_DECISION_DEVICE=vulkan cortiq decide cortiq-decision.cmf \
--skill banking77 -p "I still have not received my new card"
On multi-GPU hosts, also set
CORTIQ_DECISION_VULKAN_ADAPTER
to a unique part
of the GPU name. The same variables work with
cortiq serve
.
GPU guide →
Measured results
Quality: the model alone vs Jev 1.13
With abstention enabled, accepted answers are 97.24–98.70% correct ; the model answers 54.30–92.11% of requests locally, depending on the task. Accuracy and coverage together →
Laya: 77 choices on the same Mac
Trained CMF skill: 93.34%, 3.10 ms p50. Laya's base English checkpoint, all 77 label names at once: 35.65%, 917.59 ms p50. All 3,080 rows, both local, no oracle. A 77-option stress test, not Laya's best achievable result: shortlisting and multilingual were not tested, and training conditions differ. Full protocol, failed long-rubric run and raw results →
Speed: now on Metal and Vulkan
1.10–1.24 ms on RTX PRO 4000; 2.28–2.58 ms on Apple M4. The full path is accelerated, not just the reconstruction kernel. All 10,554 test examples keep their CPU decisions and abstentions. No retraining.
M4 uses alternating CPU/GPU requests; RTX uses a warmed continuous stream. Sparse RTX traffic is slower; Metal can have worse tails under desktop load. Protocol, tail latency and memory → · Earlier API comparison with Jev →
Cost: cloud only when needed
Local-only decisions incur $0 in API fees . The optional hybrid run used DeepSeek V4.1 Flash for hard cases, reaching 93.93% / 97.47% / 88.00% overall accuracy. Fees are extrapolated from recorded runs, not a hosting quote. Oracle setup →
Why CMF?
- One deployable file. Encoder, skills and decision rules travel together.
- Resonance, not text generation. Each label tries to reconstruct the input signal; the smallest error wins. A calibrated gate decides whether to answer.
- Your tasks, your control. Add skills from examples; keep existing skill parameters unchanged. Version and roll back learned updates.
- Local by default. Enable an external oracle only when you need one.
Add your own skill
Prepare labeled examples in
train.jsonl
and describe the labels in
question.json
:
cortiq decision add-skill cortiq-decision.cmf --skill support \
--train train.jsonl --question question.json -o support.cmf
cortiq decide support.cmf --skill support -p "Please cancel my subscription"
Example data, calibration and training options →
To serve over HTTP:
cortiq serve cortiq-decision.cmf
.
Requests, keys and router compatibility →
Benchmark scope: reused public datasets; CMF trained on train + dev, Jev given two examples per label. The model combines a Cortiq NVG-modified text encoder with trained resonance topologies. Method, data and licenses · Chart data
Resonance Routing — US 19/452,440.
Runs of infosave cmf-decision on huggingface.co
137
Total runs
0
24-hour runs
121
3-day runs
121
7-day runs
121
30-day runs
More Information About cmf-decision huggingface.co Model
Url of cmf-decision
cmf-decision huggingface.co Url
Provider of cmf-decision huggingface.co
Other API from infosave
Updated:August 22 2026
Total runs:
1.8K
Run Growth:
1.1K
Growth Rate:
61.64%
Total runs:
1.6K
Run Growth:
868
Growth Rate:
59.41%
Total runs:
849
Run Growth:
-170
Growth Rate:
-21.46%
Total runs:
534
Run Growth:
339
Growth Rate:
67.53%
Total runs:
532
Run Growth:
281
Growth Rate:
57.00%
Total runs:
229
Run Growth:
168
Growth Rate:
94.92%
Total runs:
160
Run Growth:
111
Growth Rate:
86.72%
Total runs:
94
Run Growth:
-157
Growth Rate:
-201.28%
Total runs:
77
Run Growth:
53
Growth Rate:
100.00%
Total runs:
71
Run Growth:
-144
Growth Rate:
-288.00%
Total runs:
68
Run Growth:
27
Growth Rate:
50.94%
Total runs:
61
Run Growth:
-54
Growth Rate:
-117.39%
Total runs:
66
Run Growth:
-195
Growth Rate:
-375.00%
Total runs:
26
Run Growth:
-28
Growth Rate:
-107.69%
Total runs:
20
Run Growth:
19
Growth Rate:
100.00%
Total runs:
15
Run Growth:
0
Growth Rate:
0.00%
Total runs:
12
Run Growth:
3
Growth Rate:
25.00%
Total runs:
0
Run Growth:
0
Growth Rate:
0.00%
Total runs:
0
Run Growth:
0
Growth Rate:
0.00%