seanpoyner/smolcode-coder-js-3b-tools

TEXT GENERATIONConcurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 14, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The seanpoyner/smolcode-coder-js-3b-tools is a 3.1 billion parameter LoRA fine-tune of Qwen2.5-Coder-1.5B-Instruct, developed by seanpoyner. This model is specifically trained to emit native function calls, enabling agentic coding loops for small language models. It excels at driving agentic coding assistants by correctly parsing tool calls, addressing a common limitation in smaller coder models.

Loading preview...

Overview

This model, seanpoyner/smolcode-coder-js-3b-tools, is a LoRA fine-tune of the Qwen2.5-Coder-1.5B-Instruct base model, developed by seanpoyner. Its primary purpose is to enable small language models (SLMs) to effectively drive agentic coding loops by correctly emitting native <tool_call> function calls. This addresses a critical gap where many small coder models describe tool calls as plain-text JSON, which breaks agentic tool-use runtimes.

Key Capabilities

  • Native Tool Call Emission: Fine-tuned to produce <tool_call> formatted output, compatible with runtimes like Ollama and llama.cpp.
  • Agentic Coding Support: Designed to power agentic coding assistants, specifically built for the smolcode project.
  • Optimized for Small Models: Closes the tool-use gap for tiny (≤2B parameter) models, making them more functional in agentic workflows.

Training Details

  • Base Model: Qwen/Qwen2.5-Coder-1.5B-Instruct.
  • Methodology: bf16 LoRA (r=16, α=32) applied to attention and MLP projections, utilizing assistant-only loss focusing on tool calls and final answers.
  • Data: Trained on a combination of NousResearch/hermes-function-calling-v1 and synthetic smolcode tool-use trajectories, ensuring breadth and sharpness on specific tools. The training data was rendered using the exact apply_chat_template(tools=...) used during inference to ensure byte-identical targets.
  • Version 2 Fixes: This v2 release resolves a train/inference template mismatch present in v1, ensuring high free-generation tool-call parse-rate accuracy (≥90% on held-out prompts).

Good For

  • Developers building agentic coding assistants that require precise tool call generation from small models.
  • Use cases where efficient and accurate function calling is crucial for automating coding tasks.
  • Integrating tool-use capabilities into resource-constrained environments using Tiny-Titan-class models.

Finetunes
1 models