Mima 1.5 Pico 15M
an ultra-lightweight Turkish conversational LLM
2 min readupdated
Mima 1.5 Pico 15M is a compact 15.54M-parameter large language model (LLM) fine-tuned for Turkish dialogue, multi-turn conversations, and basic tool calling. Built on the previous Mima 1 Pico architecture, it uses an 8,192-token BPE tokenizer and a 512-token context window. Thanks to its tiny footprint, it can run locally or directly in the browser via a 37.53 MB WASM build. This release is aimed at developers building lightweight Turkish chatbots or embedded assistants with minimal hardware requirements.
Key features
- Final chat release: optimized for multi-turn dialogue and accurate Turkish responses.
- Lightweight tool calling: supports a custom
@arac:syntax for triggering external tools (such as weather, a calculator, and more). - In-browser inference: available in FP16-ONNX format to run in web interfaces via WASM.
- Open data: trained on an open SFT dataset and evaluated on a dedicated micro-benchmark.
What you can do
- Build a chatbot: deploy a fast assistant in a Telegram bot or web app for simple text tasks.
- Integrate tool calling: set up automatic utility triggers (like a calculator or weather lookup) using the
@arac:tag. - Run in the browser: test the model client-side without renting a powerful server.
Example
import torch
from transformers import AutoModelForCausalLM
from tokenizers import Tokenizer
from huggingface_hub import hf_hub_download
repo = "root-sys-n/Mima-1.5-Pico-15m"
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True).eval()
tok = Tokenizer.from_file(hf_hub_download(repo, "tokenizer.json"))
prompt = "Adın ne senin?"
ids = tok.encode(prompt, add_special_tokens=False).ids
x = torch.tensor([ids], dtype=torch.long)
with torch.no_grad():
out = model.generate(x, max_new_tokens=64, do_sample=True, temperature=0.3, top_k=5, eos_token_id=3)
print(tok.decode(out[0, len(ids):].tolist()).strip())Read more
DetailsBenchmarks: overall accuracy on Mima-TR-MicroBench is 37.70%, and TruthfulQA-TR MC1 is 29.99%.
Parameters: 15.54M non-embedding (18.68M total), 7 layers, hidden size 384.
Recommended generation settings: temperature = 0.30, top_k = 5, repetition_penalty = 1.00.
Limitations: due to its tiny scale, the model cannot handle complex math, coding, or academic exams.










Comments