Laghu

The model

Parameters

151.8M (tied input and output embeddings counted once)

Layout

Gemma 4 text architecture: 28 layers, width 640, 10 query and 2 key-value heads; three sliding-window layers (512 tokens) then one global layer, repeated

Initialisation

Random weights; trained from scratch, with our own 32,768-token tokenizer (shared with Gargi)

Training

40.0B token positions of pretraining at 2K (about 70B tokens read), a 1B-token stretch at 32K, then fine-tuning on 41,467 conversations

Context

16K tokens for use (stretched at 32K, but recall fails beyond about 16K)

Modes

Thinking on (a thought block before the answer) or off. We recommend off, except for short arithmetic

Prompt format

Gemma 4 chat format; sampling defaults temperature 0.7, top-p 0.95, top-k 64

Licence

Apache-2.0 for weights and code (Copyright 2026 Udgam Labs LLP); nanochat-derived files keep their MIT notice; training data keeps each source’s licence

How to run it

With transformers, in Python. Use the shipped sampling settings: greedy decoding loops more. Decode with special tokens kept, so the thinking and the answer can be told apart.

pip install "transformers>=5.18" torch

from transformers import AutoTokenizer, AutoModelForCausalLM

tok = AutoTokenizer.from_pretrained("udgamlabs/Laghu")
model = AutoModelForCausalLM.from_pretrained("udgamlabs/Laghu")

messages = [{"role": "user", "content": "Write a short poem about the sea."}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True,
                                 enable_thinking=False,  # True for short arithmetic
                                 return_tensors="pt", return_dict=True)
out = model.generate(**inputs, max_new_tokens=512, do_sample=True,
                     temperature=0.7, top_p=0.95, top_k=64)
text = tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=False)
thinking, _, answer = text.rpartition("<channel|>")
print(answer.replace("<turn|>", "").strip())

With Ollama, LM Studio, llama.cpp or MLX. These apps turn thinking on by default; the commands below turn it off for normal chat.

# Ollama
ollama run hf.co/udgamlabs/Laghu-GGUF:Q8_0 --think=false

# LM Studio: search for udgamlabs/Laghu-GGUF in the app and download Q8_0
# (API: send "reasoning_effort": "none" to turn thinking off)

# llama.cpp: an OpenAI-compatible server on localhost:8080, 16K-token context
llama-server -hf udgamlabs/Laghu-GGUF:Q8_0 --jinja --reasoning off -c 16384

# MLX on Apple silicon
pip install mlx-lm
mlx_lm.generate --model udgamlabs/Laghu-MLX --prompt "Who are you?" \
  --chat-template-config '{"enable_thinking": false}'