EganAI/qwen3.5-9b-terminal-merge
EganAI/qwen3.5-9b-terminal-merge is an experimental 9 billion parameter Qwen3.5-based language model, developed by EganAI, with a 262,144 token context length. This model is a layer-wise merge of 16 Qwen3.5-9B variants, specifically aimed at generating terminal and CLI-style commands. It features a hybrid linear and full attention architecture, optimized for command-line task generation.
Loading preview...
Overview
EganAI/qwen3.5-9b-terminal-merge is an experimental 9 billion parameter language model based on the Qwen3.5 architecture, featuring a substantial 262,144 token context length. Developed by EganAI, this model is a layer-wise merge of 16 different Qwen3.5-9B variants, with a primary focus on generating terminal and CLI-style commands. It incorporates a hybrid linear and full attention architecture, utilizing 32 layers (8 full attention + 24 linear attention).
Key Characteristics
- Architecture: Qwen3.5 (hybrid linear + full attention) with 9B parameters.
- Context Length: Supports an extensive 262,144 tokens.
- Merge Method: Utilizes a layer-wise linear merge, combining optimized weights from 16 specialized Qwen3.5-9B models.
- Optimization Target: Specifically optimized for terminal and CLI-style command generation, covering tasks like file operations, text processing, git workflows, and system administration.
Current Status and Benchmarks
While initially presented with promising results, a re-evaluation on the 60-task Modal Harbor benchmark on March 17, 2026, indicates that this v1 merge currently performs slightly below the base Qwen/Qwen3.5-9B model. The base model achieved an 88.3% pass rate (53/60 tasks), whereas EganAI/qwen3.5-9b-terminal-merge achieved an 83.3% pass rate (50/60 tasks). Therefore, it is currently considered an experimental release rather than a confirmed upgrade over the base Qwen model for general performance.
Limitations
- Does not currently outperform the base
Qwen/Qwen3.5-9Bon the latest internal benchmarks. - Optimized specifically for terminal/CLI tasks; general-purpose performance may vary.
- Requires the
--language-model-onlyflag when serving with vLLM due to its multimodal architecture, as visual capabilities were not part of the optimization target.