iapp/OpenThai-SystemOne-MLX-bf16

OpenThai-SystemOne — mlx-bf16

OpenThai-SystemOne is an open Thai + English System One decision model: one forward pass answers typed questions (choice over up to 255 options, ordinal score, yes/no noul) about a text / JSON state with calibrated probabilities, no text generation. It is a Qwen3.5-0.8B text tower (Thai continued pre-training) plus a 256-slot decision head. This repo is a quantization of v0.3 (commit f3709948).

What is quantized: the tower including the token embeddings (MLX quantizes the embedding table too). The 256-slot decision head and the per-type temperatures stay in fp32 (head.safetensors). Quantization therefore only perturbs the hidden state the head reads.

Format: MLX bf16 for Apple Silicon (mlx-lm). mlx-lm runs the tower; the included client applies the decision head on the final hidden states. Size: 1505 MB.

Measured on a MacBook Pro M3 Max: 4-bit ≈ 19 ms per 3-question Thai decision (the PyTorch model on MPS: ~150 ms).

Usage

pip install mlx-lm torch transformers safetensors pydantic
huggingface-cli download iapp/OpenThai-SystemOne-MLX-bf16 --local-dir openthai-mlx

import sys; sys.path.insert(0, "openthai-mlx")
from openthai_systemone.mlx_client import MLXSystemOneClient
c = MLXSystemOneClient("openthai-mlx")
r = c.system_one("ร้านนี้อาหารอร่อยมาก แต่รอนานเกือบชั่วโมง", {"sentiment": {"type": "choice", "instructions": "ความรู้สึก",
                 "criteria": {"บวก": None, "ลบ": None, "กลาง": None}}})
print(r.answers["sentiment"].choice, r.answers["sentiment"].probabilities)

Accuracy vs the bf16 original (same records, single option order, first 800 per set)

Macro: public 74.3 (original 74.3), Thai 80.2 (original 80.1).

subset	bf16 original	this	Δ
public 13-subset bench
aegis2 (noul)	83.2	83.2	+0.0
boolq (noul)	79.7	80.0	+0.3
civil_comments (noul)	79.0	79.0	+0.0
helpsteer2 (score)	41.6	41.6	+0.0
massive-de-DE (choice)	88.3	88.3	+0.0
massive-en-US (choice)	88.3	88.3	+0.0
multinli (choice)	89.0	88.6	-0.3
paws (noul)	94.0	94.0	+0.0
pubmedqa (choice)	64.0	64.0	+0.0
squad2 (noul)	89.3	89.3	+0.0
summeval-consistency (score)	75.0	75.7	+0.7
summeval-relevance (score)	21.7	20.8	-0.8
vitaminc-dev (choice)	72.5	72.8	+0.3
macro, public 13-subset bench	74.3	74.3	+0.0
Thai held-out / eval sets
banking77 (choice)	59.1	59.4	+0.2
contrastive_th (choice)	80.7	81.1	+0.3
contrastive_th (noul)	83.5	82.7	-0.8
contrastive_th (score)	78.6	78.6	+0.0
massive_th (choice)	90.6	90.9	+0.2
prachathai (choice)	98.3	98.3	+0.0
prachathai (noul)	93.4	93.3	-0.1
sib200_th (choice)	77.9	78.4	+0.5
wisesight (choice)	48.9	49.0	+0.1
wongnai (score)	64.5	64.6	+0.1
xlam_tools (choice)	99.4	99.4	+0.0
xnli_th (choice)	79.8	79.8	+0.0
xnli_th (noul)	86.8	86.8	+0.0
macro, Thai held-out / eval sets	80.1	80.2	+0.0

Notes

• Scores are single-option-order accuracy on the first 800 records of each set (scripts/06_eval.py --limit 800), the same records for the original and the quantization. score subsets report exact level accuracy.
• Base model, data, training and the full benchmark tables: iapp/OpenThai-SystemOne.
• License Apache-2.0 (same as the base). Built by iApp Technology / OpenThaiGPT.