ricdomolm/talkie-1930-coder

TEXT GENERATIONConcurrency Cost:1Model Size:13BQuant:FP8Ctx Length:32kPublished:May 2, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

The ricdomolm/talkie-1930-coder is a 13 billion parameter model developed by ricdomolm, fine-tuned for agentic software engineering tasks. Starting from the talkie-1930 base, it is specifically optimized for processing and generating code within the mini-swe-agent interaction format. This model demonstrates a pass@1 score of 4.48% ± 0.69 pp on the SWE-bench-Verified-Working-Harbor benchmark, making it suitable for automated code resolution and software development workflows. Its 32768-token context length supports complex coding challenges.

Loading preview...

ricdomolm/talkie-1930-coder: Agentic Software Engineering Model

This model, developed by ricdomolm, is a 13 billion parameter language model fine-tuned for agentic software engineering tasks. It is built upon the talkie-1930 base model and specifically optimized for the mini-swe-agent interaction format.

Key Capabilities & Performance

  • Agentic Code Resolution: Fine-tuned on agentic software-engineering trajectories from the SWE-smith dataset.
  • SWE-bench Performance: Achieves a pass@1 score of 4.48% ± 0.69 pp on the SWE-bench-Verified-Working-Harbor benchmark over 5 independent evaluation runs.
  • Context Length: Supports a substantial 32768-token context window, enabling it to handle complex codebases and problem descriptions.
  • Training Details: Trained using TRL SFTTrainer with adamw_torch_fused optimizer, bf16 precision, and a maximum sequence length of 65,536, utilizing completion_only_loss.

Usage & Integration

  • Requires trust_remote_code=True for loading due to custom modeling code (modeling_talkie.py, configuration_talkie.py).
  • Designed for integration with vLLM for serving and mini-swe-agent for driving agentic evaluations.

Companion Model

  • A companion model, ricdomolm/talkie-web-coder, exists, which uses the same training recipe but starts from a web-style data pre-trained base model, achieving 5.75% ± 1.04 pp on the same evaluation.