The flower photo by Sergey Shmidt on Unsplash

İvmeLabs

Welcome to İvmeLabs! We are a Turkish HuggingFace organization that builds AI models that are small and push above their weight.

Our main proposition for our language models is that they're small, but they do one thing (or multiple things if we are crazy) right and do it well.

Our Model Families

İvme-Conversate (Codenamed Apple): İvme-Conversate is our series of language models designed for prediction and conversation using human language.

İvme-Classify (Upcoming, Codenamed Orange): İvme-Classify is our series of language models designed for zero-shot classification at small sizes.

İvme-Polisher (Upcoming, Likely Delayed, Codenamed Lime): İvme-Polisher is our series of diffusion language models (That might change; we don't know) designed for polishing user text on the go, such as fixing grammar or adding markdown styling.

İvme-Corrector (Upcoming, Codenamed Lemon): İvme-Corrector is our series of language models designed for autocorrect tasks and word prediction specifically for Turkish and its morphology.

İvme-Speak (Files Lost, Upcoming Later, Codenamed Blueberry): İvme-Speak is our series of TTS models designed to give realistic-sounding voices at small parameter sizes. Engineered for assistants and ect where you will be hearing it constantly.

İvme-CaaLM/İvme-LaaLM (Proposed, Unknown): While we can't verify it happening, we have plans to merge our sister organizations LaaLM and CaaLM into the İvme family.

İvmecikler (Proposed, Unknown): İvmecikler is our proposed family of models designed for trying to stay useful or even vaguely coherent at extreme sizes (Think sub-5M or even sub-1M). We can't verify they'll come, but they will probably, as the price of them will be dirt cheap basically.

İvme-Ultra (Proposed, Likely Upcoming, Codenamed Belladonna): İvme-Ultra is our planned first model series with MoE. The first model is going to be 1.2B, which is not really small,l but we are going to give it a weird quirk which might compensate for that. We don't want to give more details about it, as it's early, and we don't want to leak the ideas on it.

Definitely Planned Models

İvme-Conversate-v2: This will be our next model for the İvme-Conversate family. We will be experimenting with some parts of the model. But let's not spoil it too much.

İvme-Classify-v1: While the parameter count and everything overall are unknown, we are definitely going to release it sometime soon.

İvme-Corrector-v1: Likely going to be our next model. Not verified but planned to be a CNN+Transformer model.

Note the Budget

We will TRY to bring models on time, but we CAN'T really verify that they'll come as soon as we say they are coming. As we hit budget problems and noisy hosts. Even backing up İvme-Conversate-v1-Base's training data was hard for us after our host refused to boot. But don't worry, we will try our best as long as our host won't fail and our budget won't hit its limit!

Current Status

Last Updated: July 3, 2026

We finished all preparations for İvme-Conversate-v2. We are expecting to train on a TPU v6e-1 and start training tomorrow,w and probably end training and even release tomorrow.