iapp/OpenThai-SystemOne-MLX-nvfp4

OpenThai-SystemOne — mlx-nvfp4

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 NVFP4 for Apple Silicon (mlx-lm). mlx-lm runs the tower; the included client applies the decision head on the final hidden states. Size: 424 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-nvfp4 --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 73.2 (original 74.3), Thai 79.4 (original 80.1).

subset	bf16 original	this	Δ
public 13-subset bench
aegis2 (noul)	83.2	80.8	-2.4
boolq (noul)	79.7	78.7	-1.0
civil_comments (noul)	79.0	81.3	+2.3
helpsteer2 (score)	41.6	42.8	+1.2
massive-de-DE (choice)	88.3	84.0	-4.3
massive-en-US (choice)	88.3	87.7	-0.6
multinli (choice)	89.0	85.6	-3.3
paws (noul)	94.0	92.8	-1.2
pubmedqa (choice)	64.0	62.4	-1.6
squad2 (noul)	89.3	86.3	-3.0
summeval-consistency (score)	75.0	79.2	+4.2
summeval-relevance (score)	21.7	20.8	-0.8
vitaminc-dev (choice)	72.5	69.8	-2.7
macro, public 13-subset bench	74.3	73.2	-1.0
Thai held-out / eval sets
banking77 (choice)	59.1	60.6	+1.5
contrastive_th (choice)	80.7	80.7	+0.0
contrastive_th (noul)	83.5	81.0	-2.4
contrastive_th (score)	78.6	75.0	-3.6
massive_th (choice)	90.6	89.8	-0.9
prachathai (choice)	98.3	98.8	+0.5
prachathai (noul)	93.4	93.9	+0.4
sib200_th (choice)	77.9	76.5	-1.5
wisesight (choice)	48.9	48.6	-0.2
wongnai (score)	64.5	64.6	+0.1
xlam_tools (choice)	99.4	99.2	-0.1
xnli_th (choice)	79.8	77.0	-2.8
xnli_th (noul)	86.8	86.1	-0.6
macro, Thai held-out / eval sets	80.1	79.4	-0.7

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.