thinker-34M

Does this question need careful thinking? thinker-34M answers quick or think for a question, so an assistant can switch a model's slower "thinking" mode on only when it helps. It is Viveka-34M fine-tuned as a classifier and runs in milliseconds on a laptop CPU, next to router-34M.

93.6% correct on 1,798 held-out questions it never saw in training (version 3, November 2026). The first version, tested on the same questions, gets 71%.

Label	Typical questions
quick	small talk, a fact ("What is the capital of France?"), a simple explanation, a one-step sum
think	word problems, competition maths, proofs, multi-step puzzles, hard science questions

Use

from transformers import pipeline
clf = pipeline('text-classification', model='vrishv/thinker-34M')
print(clf('What is the derivative of x squared?'))

An ONNX copy (model.onnx) is included for CPUs without PyTorch, e.g. with onnxruntime and tokenizers:

import numpy as np, onnxruntime as ort
from tokenizers import Tokenizer
from huggingface_hub import hf_hub_download
tok = Tokenizer.from_file(hf_hub_download('vrishv/thinker-34M', 'tokenizer.json'))
sess = ort.InferenceSession(hf_hub_download('vrishv/thinker-34M', 'model.onnx'))
ids = tok.encode('What is the derivative of x squared?').ids[:128]
logits = sess.run(None, {'input_ids': np.array([ids]), 'attention_mask': np.ones((1, len(ids)), dtype=np.int64)})[0]
print(['quick', 'think'][int(logits.argmax())])

Training

Version 3 is built on the IAM-free Viveka-34M (v3) and was fine-tuned on a Kaggle T4 GPU (learning rate 1e-4, 3 epochs, the better of two settings kept) on 9,000 quick and 9,000 think questions. The two labels are matched length by length (and chatty lead-ins appear equally in both), so the model cannot just count words: a guess from length alone gets only about 52%.

Label	Examples from
quick	small talk from Google Synthetic-Persona-Chat (CC BY 4.0), facts from Natural Questions (NQ-open, CC BY-SA 3.0), AI2 ARC-Easy (CC BY-SA 4.0), Jigyasa Simple, Medium and Big (Apache 2.0), one-step sums made by a script
think	GSM8K and MATH (MIT), AQuA-RAT (Apache 2.0), proofs and theory from StackMathQA (CC BY 4.0), AI2 ARC-Challenge (CC BY-SA 4.0), multi-step puzzles and short-but-hard questions made by a script

Limitations

• Some of the training questions were made by a script, so real questions with unusual wording can be harder.
• English only. Questions are read up to 128 tokens.

About the maker

thinker-34M was built by Vrishvardhan Sharma (vrishv), a 9-year-old self-taught builder from India. His other work:

• GitaVardhan: a free, offline Bhagavad Gita app in 13 Indian languages, and a pocket-sized voice-controlled Gita device he designed, wired and coded himself.
• Sanskrit Vardhan: an app for learning Sanskrit through the Bhagavad Gita, with word-by-word meanings and step-by-step Pāṇinian derivations. Launched on 20 October 2026 by Prof. Shrinivasa Varakhedi, Vice-Chancellor of Central Sanskrit University.
• Kannada, Tamil and Telugu Vardhan: apps that teach everyday spoken Kannada, Tamil and Telugu.
• Hindi Vardhan (hindivardhan.in): a dictionary of pure, Sanskrit-rooted Hindi. Over 6,000 borrowed words in everyday Hindi, each with native alternatives checked against five classical Sanskrit dictionaries.
• AksharaVardhan: a handwriting reader that runs on an ordinary computer.
• graphics-studio: a free Space that makes and edits images, videos, music and narration.
• Viveka-34M: his small language model, trained from scratch, that this model is built on.
• Shodhana-34M: Viveka trained further to correct lines misread by a handwriting reader.
• Jigyasa Simple: 473 everyday science questions with short, plain-text answers.
• Jigyasa Medium and Jigyasa Big: 969 and 970 deeper science questions, built from his own questions, with answers of up to 7 and 10 sentences.
• router-34M: picks which of four offline models should answer a question; used next to this thinker.

He has memorised the Bhagavad Gita and Pāṇini's Aṣṭādhyāyī, won national Sanskrit recitation competitions, and speaks six languages. More at vrishvardhan.in.

CC BY-SA 4.0 © 2026 Vrishvardhan Sharma.