foksly/wmt26-constrained-submission

TEXT GENERATIONPricing:Input $0.37 / Cached $0.074 / Output $0.38Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:8kTool Calling:SupportedPublished:Aug 5, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

The foksly/wmt26-constrained-submission is an 8-billion parameter decoder-only causal language model developed by foksly. It is specifically designed for machine translation from English into Russian, Belarusian, Kazakh, and Armenian. This model excels in multilingual translation tasks, demonstrating strong performance against larger models in WMT26 General MT evaluations.

Loading preview...

Model Overview

The foksly/wmt26-constrained-submission is an approximately 8-billion parameter decoder-only causal language model developed by foksly. This model was specifically created as a constrained submission for the WMT26 General Machine Translation task. Its primary function is to translate English text into four target languages: Russian, Belarusian, Kazakh, and Armenian. The model supports a context length of 32,768 tokens.

Key Capabilities

  • Multilingual Machine Translation: Specializes in English-to-Russian, English-to-Belarusian, English-to-Kazakh, and English-to-Armenian translation.
  • Domain-Specific Translation: Supports WMT26 domain prompts including social, speech, news, and software, allowing for tailored translation outputs.
  • Custom Instructions: Users can provide custom instructions via --prompt or --prompt-file for flexible translation scenarios.

Performance Highlights

Evaluations against public models with up to 20B parameters show strong performance:

  • ChrF++: Achieves leading scores across all target languages (e.g., 62.3 for en-ru, 55.3 for en-kk), often outperforming larger models like TranslateGemma-12B and Qwen-3.5-9B.
  • WMT25 General MT: Demonstrates superior Accuracy and Fluency scores, and competitive MQM scores, particularly for en-ru, en-be, and en-kk, when evaluated with ORBIT-SC using GPT-5.4 as a single judge.

Usage

The model can be used via the provided inference.py script for plain or domain-specific translations, or integrated into applications using the Hugging Face Transformers library.