Intelligence shouldn’t
be heavy.

Our latest model has 151 million parameters and was pretrained on one trillion tokens. The models we compare it with are larger, and were trained on six to nineteen times more data.

After post-training, it scores above Gemma 3 270M on every benchmark except GPQA Diamond, where all three models are at chance. It leads LFM2.5-230M on BBH, MMLU Pro and GSM8K with chain of thought, and trails it on instruction following, code and GSM8K without chain of thought.

See the full Madhuram-v1 results

We publish our benchmarks, our methods and our failures as we go.

1. Oct 2026Madhuram-v1151M parameters and 1T tokens, with pretraining and post-training results against models trained on up to 19 times as much data.
2. Jul 2026Madhuram-v0.6: technical guideA 16K context window and full benchmark results against SmolLM2, Gemma 3 and LFM2.5.
3. Jul 2026Madhuram on deviceMobile inference benchmarks and comparisons, measured on a phone.
4. Jan 2026Madhuram-v0.5: technical guideArchitecture details, benchmarks and performance comparisons.
5. Jul 2025Building Madhuram-TranslateA translation model built on our Indic tokenizer, with minimal data and hardware.
6. Jun 2025Perseus: efficient language modellingRethinking positional embeddings to get more out of a small model.

Finetuning and on-device builds are custom engagements. Tell us about your case and we will tailor the work to it.

With Madhuram v2 we want to go further: larger models than 151M parameters, more training data, and a real step up in capability.

We are looking for investors and partners who believe capable models should run on the devices people already own. If that is you, we would like to hear from you.

[email protected] @maruthlabs on X