Adiuk/eyla-qwen3-8b-tools-v2

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 12, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Adiuk/eyla-qwen3-8b-tools-v2 is a compact, 8-billion parameter Qwen3-8B based model developed by Adioris / Eyla, specifically fine-tuned for on-device tool-calling. It excels at emitting structured tool calls from natural-language intent in English, Bangla, and Banglish, making it suitable for offline, privacy-preserving agentic workflows. This model integrates tools and procedures directly into its weights, eliminating the need for tool schemas in context during inference.

Loading preview...

Overview

Adiuk/eyla-qwen3-8b-tools-v2 is an 8-billion parameter, decoder-only LLM developed by Adioris / Eyla, built upon the Qwen3-8B base model. It is specifically designed as a compact, on-device tool-calling model, capable of interpreting natural-language intent and emitting structured tool calls. A key differentiator is its ability to compile tools and procedures directly into its weights, meaning it requires only a short system prompt and no tool schemas in context during inference.

Key Capabilities

  • Multilingual Tool-Calling: Supports tool-calling from English, Bangla (বাংলা), and Banglish (code-switched Bangla/English) requests.
  • On-Device Operation: Optimized for local execution, with a 4-bit quantized version requiring approximately 4.3 GB of memory and achieving around 5 seconds per turn on Apple Silicon.
  • Structured Tool Output: Emits tool calls in a consistent JSON format, such as <tool_call>{"name": "file_read", "arguments": {"path": "/path/to/file"}}</tool_call>.
  • Efficient Inference: By embedding tool knowledge, it streamlines the inference process for agentic tasks.

Good For

  • On-device agent brains/tool-routers: Ideal for use within an agent harness that parses and executes tool calls (e.g., file operations, web search, shell commands).
  • Offline & Privacy-Preserving Workflows: Suitable for applications where data cannot be sent to cloud services.
  • Bangla-first Products: Specifically designed to handle tool-calling from Bengali and Banglish requests, catering to the Bangladeshi market.

Finetunes
1 models