Mica v0.1 4B
A Qwen3.5-4B decision model with a merged LoRA, trained on English and Korean. It reads the state once and returns a probability per answer (yes/no, 2 to 255 options, or a 2 to 10 level score) from single-token label logits, with no generated text.
A personal release by Akiv. A merged rank-16 LoRA on all 32 layers with no added heads; several questions about one state share the prefill. Calibration is a single temperature (1.124) fitted on the maker's calibration set. Serves a Jev-compatible /v1/systemone endpoint through llama.cpp, with the maker's own GGUF builds in the same repo. The maker's run on JevBench's public items is not on that third-party board, and part of the hard tier was used during development. Weak on knowledge-heavy questions (MMLU-Pro 53.0 in the maker's run); instructions planted inside the state still sway answers; inputs over 8,192 tokens are rejected.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
At a glance
| Parameters | 4B |
| Base model | Qwen/Qwen3.5-4B |
| Maker | Akiv (akivet) |
| Released | 2026-09-25 |
| License | apache-2.0 |
| Reported accuracy | 67.4% |
| Reported latency | 54 ms p50 / 552 ms p95 on one RTX 3090, BF16 GGUF via llama.cpp, on JevBench's public items |
Get the weights
pip install systemonemodels
systemone pull akivet/mica
The files are served from the maker's Hugging Face repository, sky7350/Mica-v0.1-4B, and verified against the checksums recorded here.
Read more
This page was opened by System One for Akiv (akivet), who can claim the organisation and take it over at any time.