Lite-Coder/LiteCoder-Terminal-4b-sft

TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:Mar 30, 2026License:mitArchitecture:Transformer Open Weights Cold

LiteCoder-Terminal-4b-sft is a 4 billion parameter language model developed by Lite-Coder, fine-tuned from Qwen3-4B-Instruct-2507. It is specifically designed as a lightweight code agent, excelling in long-horizon terminal environments. The model demonstrates improved performance on Terminal Bench evaluations due to scaling up training data to over 11,000 trajectories with diverse agent scaffolds.

Loading preview...

LiteCoder-Terminal-4b-sft Overview

LiteCoder-Terminal-4b-sft is a 4 billion parameter model developed by Lite-Coder, fine-tuned from Qwen3-4B-Instruct-2507. It is part of a series of lightweight code agents, specifically optimized for tasks within terminal environments. The model's development focused on scaling up training data, utilizing the LiteCoder-Terminal-SFT dataset, which now includes 11,255 trajectories covering a broader task taxonomy and diverse agent scaffolds.

Key Capabilities & Performance

This model shows consistent improvements across various Terminal Bench evaluations compared to its preview version and the base Qwen3-4B-Instruct model. For instance, on Terminal Bench 1.0, it achieves a 14.69% pass@1 and 28.75% pass@4 using the Terminus 2 agent. On Terminal Bench Pro, it reaches a 21.5% pass@1.

What Makes It Different?

LiteCoder-Terminal-4b-sft stands out due to its specialized fine-tuning for long-horizon terminal environments and its focus on agentic capabilities for code-related tasks. Its training on a significantly expanded and diversified dataset of terminal trajectories makes it particularly adept at navigating and executing commands within these complex environments, offering a lightweight yet capable solution for code agents.