Skip to content

Saracura PT-BR 4B

FelhenDecision modelOpen weights

Saracura PT-BR 4B by Felhen is a decision model: it answers typed questions with a probability for each allowed answer instead of generating text. It scores 40.2 on the Jev Decision Index (rank 81).

Pricing

No hosted price is listed. The weights are open: the cost is your own compute.

Specifications

Provider
Felhen
Parameters
5B
Released
Oct 2026
Open weights
Yes
Question types
Choice (one of N options), Score (ordered rubric), Yes / no
Mechanism
Decoder LM with a decision readout
Backbone
Qwen/Qwen3.5-4B-Base
Jev-compatible request
Yes
Runs on CPU
No
License
apache-2.0
Hugging Face
felhen-ai/saracura-ptbr-4b

Benchmarks

Jev Decision IndexECE 0.179
40.2

Jev Decision Index: public 38.0, ECE 0.179, Brier 0.538. Lower ECE means its probabilities can be taken at face value.

Self-reported by the publisher

The model's own claims, often on its own sets or a self-run copy of a public suite. Not independently verified and never ranked alongside measured results.

PT-BR typed decisionsSelf-reportedPT-BR typed decisions, public split accuracy (felhen-ai board)
72.9
Jev Decision Index public suite (self-run)Self-reportedDecision Index public balanced skill, edition 0.3 kit (release-v3), self-run
38.0

Speed & latency

Median latency per call (Decision Index)
62 ms

Compare Saracura PT-BR 4B with…

FAQ

How much does Saracura PT-BR 4B cost?

Saracura PT-BR 4B has no hosted price listed; its weights are open, so the cost is the compute you run it on.

What questions can Saracura PT-BR 4B answer?

Saracura PT-BR 4B answers choice (one of n options), score (ordered rubric), yes / no questions, returning a probability for each allowed answer.

Is Saracura PT-BR 4B well calibrated?

On the Jev Decision Index its expected calibration error (ECE) is 0.179; below about 0.05, a stated 80% confidence is right roughly 80% of the time.

Is Saracura PT-BR 4B open weight, and can it run on a CPU?

Yes, Saracura PT-BR 4B has open weights. It needs a GPU to run at useful speed.

Data as of 2026-10-14. Source: Metatext decision-model catalogue, 2026-10-14 (Jev Decision Index, Hugging Face). Pricing is per 1M tokens (USD); confirm with the provider before production use.