MMOPD/Qwen3-1.7B-OT3-tau

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 16, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

MMOPD/Qwen3-1.7B-OT3-tau is a 2 billion parameter causal language model, fine-tuned from MMOPD/Qwen3-1.7B-OT3-2ep, which itself is based on Qwen/Qwen3-1.7B-Base. This model specializes in multi-turn, tool-calling agent behavior, building upon an initial 'thinking' capability. It was trained on 33,531 Tau2 tool-use trajectories, making it suitable for tasks requiring structured reasoning and tool interaction within a 32,768-token context window.

Loading preview...

Model Overview

MMOPD/Qwen3-1.7B-OT3-tau is a 2 billion parameter model developed by MMOPD, derived from MMOPD/Qwen3-1.7B-OT3-2ep and ultimately Qwen/Qwen3-1.7B-Base. Its core distinction lies in its specialized fine-tuning for multi-turn, tool-calling agent behavior, building upon an existing 'thinking' capability where answers begin with a <think> block. This model was trained alongside a 4B parameter counterpart (MMOPD/Qwen3-4B-OT3-tau) using identical data and hyperparameters, differing only in size.

Key Capabilities

  • Tool-Calling Agent Behavior: Specifically fine-tuned on the Tau2 tool-use trajectories from inclusionAI/AReaL-tau2-data to enable robust tool interaction.
  • Structured Reasoning: Inherits and enhances a 'thinking' mechanism, allowing the model to output reasoning steps before generating content or making tool calls.
  • Extended Context Window: Supports a context length of 32,768 tokens, crucial for complex multi-turn conversations and tool interactions.
  • Chat Template Compatibility: Maintains the native chat template of its base model, including the enable_thinking feature.

Training Details

The model underwent supervised fine-tuning (SFT) for 2 epochs on 33,531 Tau2 trajectories. The training procedure involved full-parameter SFT with completion-only loss, using a sequence length of 32,768 and bf16 mixed precision. Loss masking was applied to supervise the reasoning, content, tool calls, and the final stop token, while masking earlier turns and system policies.

Limitations

It's important to note that the model was trained on all Tau2 trajectories, including those with unsuccessful task outcomes, meaning it may imitate both successful and unsuccessful tool-use patterns. It inherits the limitations of its base model and has not undergone additional safety-tuning beyond what the base and Tau2 data provide.

Finetunes
1 models