AI Briefing
KO

Portuguese-specialized SLM, Zagreus-0.4B released

·2026.07.24 23:37

Key point

mii-llm has released a small language model with about 400 million parameters supporting Portuguese and English.

Details

Open source AI research lab mii-llm has released Zagreus-0.4B-por, a Portuguese-English bilingual model. This is a small model with approximately 400 million (400M) parameters, part of the Zagreus lineup focused on European languages.

Key Features and Performance:

  • Dataset: Pretrained from scratch using FineWeb Portuguese, FineWeb2 Portuguese, FinePDFs, StarCoderData, and more.
  • Benchmark: It scored an average of 0.3113 on the Portuguese versions of ARC, HellaSwag, and MMLU, a figure higher than Qwen3-0.6B-Base (0.2569) of the same scale.
  • Purpose: Rather than being a finished assistant, it is designed as a foundation model for instruction tuning, building domain-specific assistants, educational tools, and on-device applications.

This project was carried out with computing infrastructure support from Seeweb cloud and Regolo.ai.

open-sourceon-device-aismall-language-modelsportuguese-nlpmii-llm
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

RedditOpen-Weight CLM Released as Jev Alternative, Agent Decision-Making 4x FasterPyTorchKRTrillion Labs Releases GravityOCR, an OCR Model Applying Self-Speculative DecodingHacker NewsJevper Released: A Jev Classification Wrapper for OpenAI-Compatible Models with Automatic Logprobs DetectionGeekNewsRootless Linux Sandbox 'Drop' Released for Isolating AI Agents and Third-Party ProgramsPyTorchKRPostgreSQL Extension pgjev Released, Supporting Natural Language Filtering Without EmbeddingsTLDR AIJared Palmer Releases 'Kev', a Lightweight Decision-Making Model Based on Qwen3.5