KookiesXy/llm1b-1.5b-toolchat

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 11, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

KookiesXy/llm1b-1.5b-toolchat is a 1.54 billion parameter instruction-tuned causal language model, warm-started from Qwen2.5-1.5B-Instruct, specifically optimized for tool-calling and chat coherence. This model was trained rapidly in approximately one hour on a curated dataset, demonstrating efficient SFT for function-calling capabilities. It excels at generating precise JSON tool calls and engaging in fluent multi-sentence conversations, making it suitable for agentic applications requiring reliable tool interaction.

Loading preview...

KookiesXy/llm1b-1.5b-toolchat: A Rapidly Trained Tool-Using LLM

This model, llm1b-1.5b-toolchat, is a 1.54 billion parameter language model based on Qwen2.5-1.5B-Instruct, distinguished by its rapid and efficient training for tool-calling capabilities. It was trained in just 61 minutes on a single Modal B300 instance, demonstrating a highly optimized SFT process.

Key Capabilities

  • Tool Calling: Emits perfect <tool_call>{"name": ..., "arguments": {...}}</tool_call> JSON given relevant tool schemas, enabling robust agentic workflows. Achieved 12/12 tool tasks passed in evaluations.
  • Chat Coherence: Provides fluent multi-sentence answers, stories, haikus, explanations, and plans, with an average chat coherence score of 1.0.
  • Efficient Training: Achieved a throughput of ~48.9k train tokens/second (bf16, SDPA, grad checkpointing) over 179.5M tokens, utilizing a tightly-packed SFT pass.

Training Methodology Highlights

  • Data: Utilized a small, high-quality dataset (201,754 unique examples) combining chat data (e.g., mlabonne/orpo-dpo-mix-40k, databricks/databricks-dolly-15k, HuggingFaceH4/ultrachat_200k) and tool data (glaiveai/glaive-function-calling-v2). Tool examples constituted roughly 55% of training tokens.
  • Packing: Examples were tokenized and packed into 4,096-token blocks with an assistant-only label mask to prevent copy-collapse issues.

Good For

  • Developers seeking a compact yet capable model for function-calling and tool-use in agentic applications.
  • Use cases requiring fast inference and reliable JSON output for tool interactions.
  • Scenarios where resource-efficient training and deployment are critical.

Finetunes
1 models