Instructions to use sraivante/TinyThinkModel-72M-V1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use sraivante/TinyThinkModel-72M-V1 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="sraivante/TinyThinkModel-72M-V1")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("sraivante/TinyThinkModel-72M-V1")
model = AutoModelForCausalLM.from_pretrained("sraivante/TinyThinkModel-72M-V1", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use sraivante/TinyThinkModel-72M-V1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "sraivante/TinyThinkModel-72M-V1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "sraivante/TinyThinkModel-72M-V1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/sraivante/TinyThinkModel-72M-V1

SGLang

How to use sraivante/TinyThinkModel-72M-V1 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "sraivante/TinyThinkModel-72M-V1" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "sraivante/TinyThinkModel-72M-V1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "sraivante/TinyThinkModel-72M-V1" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "sraivante/TinyThinkModel-72M-V1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use sraivante/TinyThinkModel-72M-V1 with Docker Model Runner:
```
docker model run hf.co/sraivante/TinyThinkModel-72M-V1
```

TinyThinkModel-72M-V1 / code /common.py

sraivante

Release TinyThinkModel-72M-V1 as an owner-authorized private snapshot

fac8bb8 verified 27 days ago

Raw History Blame Contribute Delete

2.57 kB

"""Shared, deterministic model/data definitions. Copyright (c) 2026 sraivante."""
from __future__ import annotations
import hashlib
import json
import re
import unicodedata
from datetime import datetime, timezone
from pathlib import Path

SEED = 20261017
CONTEXT = 1024
VOCAB = 33069
SPECIALS = ["<|pad|>", "<|endoftext|>", "<|im_end|>", "<|im_start|>", "<think>", "</think>"]
MAX_SOURCE_BYTES = 20_000_000_000
MAX_FILE_BYTES = 7_000_000_000

def now():
    return datetime.now(timezone.utc).isoformat()

def sha(path):
    h = hashlib.sha256()
    with Path(path).open('rb') as f:
        for chunk in iter(lambda: f.read(8 * 1024 * 1024), b''):
            h.update(chunk)
    return h.hexdigest()

def dump(path, value):
    path = Path(path)
    path.parent.mkdir(parents=True, exist_ok=True)
    tmp = path.with_name(path.name + '.partial')
    tmp.write_text(json.dumps(value, indent=2, ensure_ascii=False) + '\n', encoding='utf-8')
    tmp.replace(path)

def prompt_key(prompt):
    # All identical normalized questions share a split, including different solutions.
    s = re.sub(r'\s+', ' ', unicodedata.normalize('NFKC', str(prompt))).strip().casefold()
    return hashlib.sha256(s.encode('utf-8')).digest()

def split_for(prompt):
    bucket = int.from_bytes(prompt_key(prompt)[:8], 'big') % 1000
    return 'validation' if bucket < 5 else ('test' if bucket < 10 else 'train')

def render_row(row):
    prompt = str(row.get('prompt') or '').strip()
    answer = str(row.get('answer') or '').strip()
    thought = str(row.get('think') or '').strip()
    if not prompt or not (answer or thought):
        return None
    # Preserve source Unicode, math, code indentation and newlines. No ASCII filtering.
    reasoning = '<think>\n' + thought + '\n</think>\n' if thought else ''
    return ('<|im_start|>user\n' + prompt + '<|im_end|>\n'
            + '<|im_start|>assistant\n' + reasoning + answer + '<|im_end|>\n<|endoftext|>')

def model_config():
    from transformers import Qwen3Config
    return Qwen3Config(
        vocab_size=VOCAB, hidden_size=512, intermediate_size=1536,
        num_hidden_layers=14, num_attention_heads=8, num_key_value_heads=4,
        head_dim=128, max_position_embeddings=CONTEXT,
        hidden_act='silu', rms_norm_eps=1e-6, rope_theta=1_000_000.0,
        attention_bias=False, attention_dropout=0.0, tie_word_embeddings=True,
        use_sliding_window=False, sliding_window=None, max_window_layers=14,
        pad_token_id=0, bos_token_id=1, eos_token_id=2, use_cache=False,
        architectures=['Qwen3ForCausalLM'],
    )

	"""Shared, deterministic model/data definitions. Copyright (c) 2026 sraivante."""
	from __future__ import annotations
	import hashlib
	import json
	import re
	import unicodedata
	from datetime import datetime, timezone
	from pathlib import Path

	SEED = 20261017
	CONTEXT = 1024
	VOCAB = 33069
	SPECIALS = ["<\|pad\|>", "<\|endoftext\|>", "<\|im_end\|>", "<\|im_start\|>", "<think>", "</think>"]
	MAX_SOURCE_BYTES = 20_000_000_000
	MAX_FILE_BYTES = 7_000_000_000

	def now():
	return datetime.now(timezone.utc).isoformat()

	def sha(path):
	h = hashlib.sha256()
	with Path(path).open('rb') as f:
	for chunk in iter(lambda: f.read(8 * 1024 * 1024), b''):
	h.update(chunk)
	return h.hexdigest()

	def dump(path, value):
	path = Path(path)
	path.parent.mkdir(parents=True, exist_ok=True)
	tmp = path.with_name(path.name + '.partial')
	tmp.write_text(json.dumps(value, indent=2, ensure_ascii=False) + '\n', encoding='utf-8')
	tmp.replace(path)

	def prompt_key(prompt):
	# All identical normalized questions share a split, including different solutions.
	s = re.sub(r'\s+', ' ', unicodedata.normalize('NFKC', str(prompt))).strip().casefold()
	return hashlib.sha256(s.encode('utf-8')).digest()

	def split_for(prompt):
	bucket = int.from_bytes(prompt_key(prompt)[:8], 'big') % 1000
	return 'validation' if bucket < 5 else ('test' if bucket < 10 else 'train')

	def render_row(row):
	prompt = str(row.get('prompt') or '').strip()
	answer = str(row.get('answer') or '').strip()
	thought = str(row.get('think') or '').strip()
	if not prompt or not (answer or thought):
	return None
	# Preserve source Unicode, math, code indentation and newlines. No ASCII filtering.
	reasoning = '<think>\n' + thought + '\n</think>\n' if thought else ''
	return ('<\|im_start\|>user\n' + prompt + '<\|im_end\|>\n'
	+ '<\|im_start\|>assistant\n' + reasoning + answer + '<\|im_end\|>\n<\|endoftext\|>')

	def model_config():
	from transformers import Qwen3Config
	return Qwen3Config(
	vocab_size=VOCAB, hidden_size=512, intermediate_size=1536,
	num_hidden_layers=14, num_attention_heads=8, num_key_value_heads=4,
	head_dim=128, max_position_embeddings=CONTEXT,
	hidden_act='silu', rms_norm_eps=1e-6, rope_theta=1_000_000.0,
	attention_bias=False, attention_dropout=0.0, tie_word_embeddings=True,
	use_sliding_window=False, sliding_window=None, max_window_layers=14,
	pad_token_id=0, bos_token_id=1, eos_token_id=2, use_cache=False,
	architectures=['Qwen3ForCausalLM'],
	)