Arthur-75/storm-qwen3-0.6B

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 4, 2026Architecture:Transformer0.0K Featherless Exclusive Cold

Arthur-75/storm-qwen3-0.6B is an 0.8 billion parameter language model based on the Qwen3 architecture, specifically designed for self-supervised lexical query expansion. It utilizes a novel Stepwise Token Optimization with Reward-guided beaM search (STORM) framework to generate retrieval-effective keywords for search systems. This model excels at enhancing lexical retrievers like BM25 by overcoming vocabulary mismatch, offering a competitive and infrastructure-light alternative to dense neural retrieval, and demonstrates strong zero-shot transfer capabilities across 18 languages.

Loading preview...

Overview

Arthur-75/storm-qwen3-0.6B is an 0.8 billion parameter Qwen3-based language model developed within the STORM (Stepwise Token Optimization with Reward-guided beaM search) framework. This model is specifically engineered for lexical query expansion, aiming to improve the performance of traditional lexical retrievers like BM25 by generating highly relevant keywords. It addresses the limitations of vocabulary mismatch in standard lexical search and the challenges of training effective LLM rewriters.

Key Capabilities

  • Self-supervised Keyword Generation: Employs a unique self-supervised approach where generation is guided by retrieval metrics, turning retrieval rewards into a token-level signal for effective vocabulary exploration.
  • Enhanced Lexical Retrieval: Significantly boosts the performance of lexical search systems, matching or surpassing competitive LLM rewriters and larger proprietary models.
  • Infrastructure-Light: Provides an efficient alternative to dense neural retrieval, operating on standard inverted indexes without requiring frequent rebuilding as models evolve.
  • Multilingual Transfer: Demonstrates strong zero-shot transfer capabilities across 18 languages (MIRACL benchmark), outperforming dedicated multilingual dense retrievers.

Good For

  • Improving Search Relevance: Ideal for applications requiring enhanced search accuracy and recall using lexical methods.
  • Query Rewriting and Expansion: Suitable for systems that need to expand user queries with semantically related and retrieval-effective keywords.
  • Efficient Retrieval Systems: Excellent for scenarios where infrastructure simplicity and fast retrieval (comparable to plain BM25) are critical.
  • Multilingual Search: Highly effective for cross-lingual information retrieval tasks without requiring language-specific fine-tuning.