surogate/Surogate-3.5-4B

VISIONConcurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 4, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Surogate-3.5-4B is a 4.5 billion parameter bilingual (Romanian/English) assistant model developed by Invergent, featuring a dual-mode reasoner that performs step-by-step reasoning in the prompt's language. It is optimized for clean Romanian text generation, achieving significantly lower rates of invented word-forms, English leakage, and missing diacritics compared to general-purpose models. This model excels in instruction following and mathematical reasoning, making it suitable for applications requiring precise bilingual interaction and robust problem-solving.

Loading preview...

Surogate-3.5-4B: A Romanian-First Bilingual Assistant

Surogate-3.5-4B is a 4.5 billion parameter dual-mode reasoning model developed by Invergent, designed for both Romanian and English language tasks. It stands out for its exceptional Romanian text quality, demonstrating significantly cleaner output with minimal invented word-forms, English leakage, and missing diacritics compared to other models of its size.

Key Capabilities

  • Dual-Mode Reasoning: The model can perform step-by-step reasoning in the language of the prompt (Romanian or English) when thinking is enabled, or provide direct answers when disabled, with no leakage between modes.
  • Superior Romanian Text Quality: Achieves 0.99 invented word-forms per 1k words, 0.70 English leakage per 1k, and 0.23 missing diacritics per 1k, based on a 13,963-prompt benchmark.
  • Strong Performance: Demonstrates competitive scores across various benchmarks, including MMLU-Pro (76.7), GPQA Diamond (73.9), and robust instruction following (IFEval 74.1). It also shows strong mathematical reasoning in both Romanian (RO GSM8K 0.712) and English (EN GSM8K 0.880).
  • Multilingual Support: Optimized for Romanian-first workflows while fully supporting English, with solid translation capabilities (chrF2 scores of 53.2 EN→RO and 62.6 RO→EN).

Good For

  • Applications requiring high-quality, clean Romanian text generation.
  • Bilingual (Romanian/English) assistant tasks where precise language and reasoning are critical.
  • Use cases demanding step-by-step problem-solving in the user's prompt language.
  • Instruction following and mathematical reasoning tasks in a bilingual context.

Finetunes
1 models