ajvikram/toolcall-2b

VISIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2.3BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 9, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

ajvikram/toolcall-2b is a 2.3 billion parameter function-calling model, fine-tuned from Qwen/Qwen3.5-2B, specifically designed for local agent tool routing. It excels at determining which tool to call and with what arguments, significantly outperforming its base model on the Berkeley Function Calling Leaderboard v4 harness for single-turn and multi-turn scenarios. This model is optimized for efficient and accurate tool selection in agentic workflows, offering a cost-effective alternative to larger frontier models for tool-use decisions.

Loading preview...

Overview

ajvikram/toolcall-2b is a 2.3 billion parameter model, fine-tuned from Qwen/Qwen3.5-2B, specialized in function calling for local agent tool routing. It is engineered to efficiently decide which tool to invoke and with what parameters, a critical and frequent decision in agentic tasks. The model was trained using supervised fine-tuning with 16-bit LoRA on a diverse dataset including ToolACE, Hermes function-calling v1, Glaive v2, and custom-generated stateful simulator episodes.

Key Capabilities

  • Enhanced Function Calling: Achieves a score of 36.35 overall on the Berkeley Function Calling Leaderboard v4 harness, significantly improving upon the base Qwen3.5-2B's 33.85. The 20-category mean (excluding web search) is 59.05 compared to the base's 48.13.
  • Strong Single-Turn Performance: Scores 83.25 on non-live single-turn tasks and 74.32 on live single-turn tasks, showing substantial gains over the base model.
  • Improved Multi-Turn Reasoning: Demonstrates better performance in multi-turn scenarios, with all four multi-turn categories showing improvement over the base model.
  • Qwen3.5 Tool-Call Format: Retains the native Qwen3.5 tool-call format, allowing for seamless integration into existing Qwen3.5-compatible stacks.
  • Quantized Builds Available: GGUF builds (e.g., Q4_K_M, Q5_K_M, Q8_0, f16) are provided for CPU inference, enabling local deployment on consumer hardware.

Good For

  • Local Agent Tool Routing: Ideal for applications requiring efficient and accurate tool selection within agentic systems, especially where cost and latency are critical.
  • Function Calling Tasks: Excels in scenarios demanding precise function call generation and argument extraction.
  • Integrating with Qwen3.5 Ecosystems: Easily integrates into existing workflows that parse Qwen3.5's XML tool-call format.

Limitations

  • Refusal Behavior: The model's refusal behavior is noted to be below the base model, meaning it may be more prone to attempting requests even when no suitable tool is available.
  • English Only: Evaluated exclusively on English tool-calling tasks.

Finetunes
2 models