Marul V7 launches Turkish-first LLM

Marul V7 is small, rough, and probably not production-ready, but that is also why it is interesting: it is a from-scratch Turkish LLM effort instead of another wrapper around a global model.

• –The custom tokenizer and architecture claim matters more than the parameter count, because Turkish morphology often exposes weaknesses in generic multilingual models.
• –At 258M parameters, expectations should stay modest: basic chat, summarization, markdown, and simple coding are plausible; strong reasoning and deep knowledge are not.
• –The launch is more useful as a community feedback loop than a benchmark event, especially since public evals, weights, datasets, and training details are not yet provided.
• –The web app also lists Llama 3 70B alongside Marul V7, so users should be careful about which model they are actually testing.