Tamil-first LLM Agaram scales to 2.5B after failed 1.5B attempt, open weights next week
Mechramc · x · 2026-10-10
Indie developer Mechramc shares an update on Agaram, a Tamil-first model trained from scratch at Murai Labs, scaling from 1.5B to 2.5B. The 1.5B version failed — it couldn't stop generating and looped responses — pushing the author into a full rework: studying more papers, rebuilding the corpus, and doubling synthetic data.
- First-of-its-size Tamil-first model pretrained from scratch on 50B tokens with its own tokenizer
- Supports Tamil + Tanglish chat and tool calling; pretraining and chat fine-tuning done
- Final training round on identity, tool use, and safety is underway
- Results and open weights promised next week if they hold up
Related event: Indie Tamil-First Model Scales to 2.5B After Openly Admitting 1.5B Failure(3 posts)→
More from Research
- EMNLP paper: diffusion LMs internally encode latent denoising timestep representations — gsarti_ · 2026-10-23
- KU Leuven and VIB open two Group Leader professorships in AI for Biology, start 2027 — jpillowtime · 2026-10-23
- Subspace Tuning curbs LLM catastrophic forgetting by freezing principal singular directions — scottniekum · 2026-10-23
- Perceptic launches PercepticOS, an AI workspace for 'scientific superintelligence' — nathanbenaich · 2026-10-23
- EPFL's Bosselut to talk long context as test-time learning at ELLIS seminar — andre_t_martins · 2026-10-23
- DRATS: Adaptive Task Sampling Boosts Worst-Task Performance in Multi-Task RL — Cohere · 2026-10-23