system-one-minicpm5-2b-q8 — quality-tier System One model (8-bit)

MiniCPM5-2B-Base (Apache-2.0) fine-tuned with the system-one frozen recipe, LoRA-fused and quantized to 8 bits with MLX (8.501 bits/weight, 2.5 GB). A System One decision model: typed questions (choice / score / noul) over a JSON state, answered with calibrated probabilities from one forward pass — no generation. Part of system-one, an open learning project rebuilding the behavior of Typesafe's Jev. Not affiliated with or endorsed by TypeSafe.

Files

• fused, 8-bit quantized model (MLX format; load with --model <dir>, no adapter)
• temperature.json — per-primitive temperatures (choice 0.961 / noul 1.073 / score 1.387), fitted under this exact quantized identity and bound to these weights — the serving engine fails closed on any mismatch.
• the unquantized LoRA adapter is published alongside for retraining/re-fusing.

Training and quantization

Same frozen recipe as the volume tier (public recasts + synthetic scenarios teacher-scored by pinned jev-1.13.0 — disclosed deliberately — with overconfidence filtering; readout CE + RPS loss, LR 1e-5, seed 42). Quantization was gate-checked: zero argmax flips across the drift battery, dev quality within noise of BF16, median 1.61x serving speedup. 4-bit was measured and rejected (quality loss, no speed gain over 8-bit on Apple Silicon).

Evaluation (full protocol and CIs in the repo)

Split	Accuracy	ECE	Notes
Development (n=1,128, calibrated)	86.6%	0.019	used for steering; optimistic
Fresh out-of-family test (n=1,048, spent once)	73.3%	0.080	+6.8 pts over the 0.6B tier, CI [+4.5, +9.2]; NLL/Brier also significantly better

Per primitive (dev): choice 94.2 / noul 90.0 / score 65.9. Confident errors at t>=0.95: 1.2% in-family.

Shift warning (measured): in-family thresholds do not transfer — out-of-family confident errors reached ~14% at t>=0.9. Use the runtime's per-workload temperature fitting (POST /v1/calibrations, ~100 labeled examples) before trusting thresholds on a new workload; it repaired scalar miscalibration to 3-4% confident errors in the repo's resampled experiment.

Use

git clone https://github.com/mpuig/system-one && cd system-one && uv sync
hf download mpuig/system-one-minicpm5-2b-q8 --local-dir system-one-minicpm5-2b-q8
uv run python scripts/serve.py --model system-one-minicpm5-2b-q8 \
  --temperature system-one-minicpm5-2b-q8/temperature.json

Light 3-question request: ~195 ms p50 on an M4 Max. Apple Silicon required (MLX). Research prototype, not a certified production service; the honest failure record lives in the repo's decision log and experiments file.