openthai/ openthai-r1-32b:q4_K_M

2 Downloads Updated 15 hours ago

Thai reasoning model that shows its step-by-step thinking — beats DeepSeek R1 70B and Typhoon R1 70B on Thai benchmarks (avg 71.58 vs 63.31/65.42) at half their size. ~24 GB RAM.

ollama run openthai/openthai-r1-32b:q4_K_M

curl http://localhost:11434/api/chat \
  -d '{
    "model": "openthai/openthai-r1-32b:q4_K_M",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='openthai/openthai-r1-32b:q4_K_M',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'openthai/openthai-r1-32b:q4_K_M',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 15 hours ago

15 hours ago

0a038a56ad52 · 20GB ·

model

archqwen2

·

parameters32.8B

·

quantizationQ4_K_M

20GB

template

{{ if .System }}{{ .System }}{{ end }}{{ if .Prompt }}<｜User｜>{{ .Prompt }}{{ end }}<｜Assistan

147B

license

Available for Research and Commercial use. Licence metadata on the source repo is other/qwen. See ht

371B

params

{ "num_ctx": 32768, "num_predict": 16384, "stop": [ "<｜User｜>", "<�

184B

Readme

OpenThai R1 32B 🇹🇭

A Thai reasoning model that shows its work — and beats 70B models at half their size.

ollama run openthai/openthai-r1-32b

R1 thinks step by step through mathematics, logic and code — in Thai. On a seven-benchmark average it scores 71.58 vs 63.31 (DeepSeek R1 Distill 70B) and 65.42 (Typhoon R1 Distill 70B), while fitting on hardware half the size.

Benchmarks (SkyThought, one identical protocol)

Benchmark	OpenThai R1 32B	DeepSeek R1 70B	Typhoon R1 70B
AIME24-TH	56.67	33.33	53.33
MATH500-TH	83.80	75.40	81.00
LiveCodeBench-TH	62.16	53.15	47.75
OpenThaiEval	76.05	74.17	77.59
Average (7 tasks)	71.58	63.31	65.42

Full table: Thai LLM Leaderboard · Technical report (arXiv:2504.01789)

Requirements & tips

~24 GB RAM/VRAM (Q4_K_M)
Reasoning traces are long — give it output room
Best at analytical tasks; use openthai-1.6-72b for general chat
OpenAI-compatible API at http://localhost:11434/v1

Links

Website · GGUF weights (Q4/Q5/Q8) · Discord

Free and open-weight, from the OpenThai project (formerly OpenThaiGPT) — AIEAT × iApp Technology.

# OpenThai R1 32B 🇹🇭

**A Thai reasoning model that shows its work — and beats 70B models at half their size.**

```
ollama run openthai/openthai-r1-32b
```

R1 thinks step by step through mathematics, logic and code — in Thai. On a seven-benchmark
average it scores **71.58 vs 63.31 (DeepSeek R1 Distill 70B) and 65.42 (Typhoon R1 Distill 70B)**,
while fitting on hardware half the size.

## Benchmarks (SkyThought, one identical protocol)

| Benchmark | **OpenThai R1 32B** | DeepSeek R1 70B | Typhoon R1 70B |
|---|---|---|---|
| AIME24-TH | **56.67** | 33.33 | 53.33 |
| MATH500-TH | **83.80** | 75.40 | 81.00 |
| LiveCodeBench-TH | **62.16** | 53.15 | 47.75 |
| OpenThaiEval | 76.05 | 74.17 | **77.59** |
| **Average (7 tasks)** | **71.58** | 63.31 | 65.42 |

Full table: [Thai LLM Leaderboard](https://huggingface.co/spaces/openthaigpt/thai-llm-leaderboard) ·
[Technical report (arXiv:2504.01789)](https://arxiv.org/abs/2504.01789)

## Requirements & tips

- **~24 GB RAM/VRAM** (Q4_K_M)
- Reasoning traces are long — give it output room
- Best at analytical tasks; use [openthai-1.6-72b](https://ollama.com/openthai/openthai-1.6-72b) for general chat
- OpenAI-compatible API at `http://localhost:11434/v1`

## Links

[Website](https://openthai.aieat.or.th/ollama) ·
[GGUF weights (Q4/Q5/Q8)](https://huggingface.co/openthaigpt/openthaigpt-r1-32b-instruct-GGUF) ·
[Discord](https://discord.gg/7KDdKkBGUs)

*Free and open-weight, from the OpenThai project (formerly OpenThaiGPT) — AIEAT × iApp Technology.*

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)