OpenThai-SystemOne — NVFP4A16

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 Linear layers of the tower. The token embeddings, the 256-slot decision head and the per-type temperatures stay in bf16. Quantization therefore only perturbs the hidden state the head reads.

Format: NVFP4 weight-only (FP4 weights, 16-bit activations). compressed-tensors (llm-compressor) checkpoint. Loaded through transformers the weights are decompressed to bf16 at load time (same speed as bf16, smaller download); native FP8 / INT8 / FP4 kernels need a runtime with this architecture (the decision head is custom, so vLLM does not serve it out of the box). NVFP4 kernels need NVIDIA Blackwell.

Size: 790 MB (bf16 original: 1,509 MB).

Usage

pip install torch transformers safetensors pydantic && pip install compressed-tensors

from transformers import AutoModel, AutoTokenizer        # trust_remote_code files are in this repo
model = AutoModel.from_pretrained("iapp/OpenThai-SystemOne-NVFP4A16", trust_remote_code=True)
# or with the pip package (git+https://github.com/iapp-technology/openthai-systemone):
from openthai_systemone import SystemOneClient
c = SystemOneClient("iapp/OpenThai-SystemOne-NVFP4A16")
r = c.system_one("ร้านนี้อาหารอร่อยมาก แต่รอนานเกือบชั่วโมง", {"sentiment": {"type": "choice", "instructions": "ความรู้สึก",
                 "criteria": {"บวก": None, "ลบ": None, "กลาง": None}}})

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

Macro: public 73.6 (original 74.3), Thai 79.8 (original 80.1).

subset	bf16 original	this	Δ
public 13-subset bench
aegis2 (noul)	83.2	82.8	-0.4
boolq (noul)	79.7	78.3	-1.3
civil_comments (noul)	79.0	76.0	-3.0
helpsteer2 (score)	41.6	41.6	+0.0
massive-de-DE (choice)	88.3	86.3	-2.0
massive-en-US (choice)	88.3	88.6	+0.3
multinli (choice)	89.0	86.6	-2.3
paws (noul)	94.0	92.4	-1.6
pubmedqa (choice)	64.0	62.0	-2.0
squad2 (noul)	89.3	87.0	-2.3
summeval-consistency (score)	75.0	78.5	+3.5
summeval-relevance (score)	21.7	25.0	+3.3
vitaminc-dev (choice)	72.5	71.6	-0.8
macro, public 13-subset bench	74.3	73.6	-0.7
Thai held-out / eval sets
banking77 (choice)	59.1	59.0	-0.1
contrastive_th (choice)	80.7	79.1	-1.7
contrastive_th (noul)	83.5	81.9	-1.6
contrastive_th (score)	78.6	80.4	+1.8
massive_th (choice)	90.6	89.5	-1.1
prachathai (choice)	98.3	98.3	+0.0
prachathai (noul)	93.4	93.7	+0.3
sib200_th (choice)	77.9	77.9	+0.0
wisesight (choice)	48.9	49.6	+0.8
wongnai (score)	64.5	64.9	+0.4
xlam_tools (choice)	99.4	99.4	+0.0
xnli_th (choice)	79.8	77.6	-2.1
xnli_th (noul)	86.8	86.4	-0.4
macro, Thai held-out / eval sets	80.1	79.8	-0.3

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.