Are the benchmark numbers real?

They come from the repo evaluation code and a local server run: perfect on the easy and original public tiers, lower on the hard tier, with per-item output in results/public231. Server and PyTorch numbers differ slightly on hard.

Can I swap this in for Jev without changing clients?

Yes for the wire format: it speaks /v1/systemone so Jev clients work unchanged. Accuracy differs from hosted Jev, especially on knowledge and long-policy items, so check the tables before swapping.

What GPU and download size should I expect?

Docker pulls the 9.7 GB BF16 GGUF on first start, or set MICA_GGUF to a smaller file like Q5_K_M for an 8 GB GPU. The image builds for CUDA arch 86, 89 and 90 by default.

Does text planted in the state still steer its answers?

Partly. A wrong-option note buried in state drops accuracy but still leaves most answers right, while a right-option note lifts it, so such notes still influence the output.

What hardware and download sizes should I expect?

With Docker the first start downloads the 9.7 GB BF16 GGUF; -e MICA_GGUF=mica-v0.1-4b-Q5_K_M.gguf uses a smaller 3.51 GB file. On an RTX 3090 the BF16 model runs at 54 ms p50 and Q4_K_M at 47 ms, and Q5_K_M suits an 8 GB GPU.

What license is it under?

Apache-2.0. The base is Qwen/Qwen3.5-4B with a merged rank-16 LoRA on all 32 layers, distributed as merged BF16 safetensors and GGUF quantizations from the sky7350/Mica-v0.1-4B repository.

How does it handle long inputs and calibration?

Inputs longer than 8,192 tokens are rejected with HTTP 400 rather than truncated, and on the held-out set ECE is 5.4% with 2.5% of answers wrong at confidence 0.9 or higher.

CommunityJev Demos & Experiments

Mica-v0.1-4B

Small decision model that returns per-answer probabilities for yes/no, 2-255 choices, or 2-10 level scores in one prefill over the SystemOne format.

See highlights

Open github.com

Stars: 36
Language: Python
License: APACHE-2.0
Published by: Community
Author: akivet
Last verified: Oct 2026

akivet/Mica-v0.1-4B

User repository on GitHub

View on GitHub

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Mica is a small decision model. You give it a state, a question and the allowed answers, and it returns a probability for each answer: yes/no, a choice among 2–255 options, or a score with 2–10 levels. It reads the input once and generates no text, so a decision costs one prefill. It speaks the TypeSafe /v1/systemone format, so clients written for Jev work unchanged. It was trained on English and Korean. Base model: Qwen/Qwen3.5-4B (revision 851bf6e) with a merged rank-16 LoRA on all 32 layers.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import

Works forJev for Classification Jev for Ranking

Used inJev for Software & DevTools

Languages: Python

Highlights

Speaks TypeSafe /v1/systemone so Jev clients work unchanged; trained on English and Korean.
Built on Qwen3.5-4B with a merged rank-16 LoRA; ships BF16 plus Q8_0 through Q4_0 GGUF files.
Reads input once with no text output, and several questions about one state share its prefill.
Rejects inputs over 8,192 tokens with HTTP 400 instead of truncating them.

Watch out

Weakest on long policy documents and knowledge-heavy questions, and planted instructions in state still move answers somewhat. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

open-jev-typed-decision-engine

Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.