Terminus-4B: Can a Smaller Model Replace Frontier LLMs at Agentic Execution Tasks?

2025-09-28 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, extended

Summary

Terminus-4B, a 4-billion-parameter small language model (SLM) based on Qwen3-4B, has been specifically post-trained to serve as an Execution Subagent for coding agents, aiming to replace expensive frontier LLMs in terminal execution tasks. Developed through a two-stage pipeline involving Supervised Finetuning (SFT) and Reinforcement Learning (RL) with a rubric-based LLM-as-judge reward, Terminus-4B significantly reduces the main agent's token usage by up to ~30% without compromising performance on benchmarks like SWE-Bench Pro and SWE-Bench C#. This specialized SLM not only matches but often surpasses the performance of frontier models such as Claude Sonnet/Opus and GPT-5.3-Codex in handling verbose terminal outputs, thereby preserving the main agent's context window and improving overall efficiency.

Key takeaway

For Machine Learning Engineers building coding agents that handle extensive terminal execution, you should consider adopting a subagent architecture with a specialized small language model like Terminus-4B. This approach can reduce your main agent's token consumption by up to 30% and maintain or even improve performance compared to using expensive frontier LLMs for subtasks. Implement a two-stage SFT and RL training pipeline to ensure your SLM effectively summarizes verbose outputs, optimizing both cost and context management.

Key insights

Finetuned small language models can match or exceed frontier LLM performance for specialized agentic subtasks.

Principles

• Subagent architectures isolate verbose outputs, preserving main agent context.
• Two-stage SFT and RL training is crucial for SLM task competence.
• Rubric-based LLM-as-judge rewards provide rich, multi-dimensional training signals.

Method

Train Terminus-4B (Qwen3-4B) via SFT on telemetry data, then RL with GRPO using a rubric-based LLM-as-judge reward against frontier LLM-generated reference trajectories.

In practice

• Delegate terminal execution to a specialized subagent to reduce token costs.
• Implement a structured summary output format for subagent responses.
• Use a two-stage SFT+RL pipeline for training specialized SLMs.

Topics

• Terminus-4B
• Coding Agents
• Subagent Architecture
• Small Language Models
• Reinforcement Learning
• Token Efficiency

Code references

• THUDM/slime

Best for: AI Architect, MLOps Engineer, CTO, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Production Sub-agents for LLM Post Training — Sub-agent architectures with Claude Code can drastically reduce ML post-training time by parallelizing tasks and mitigating context limitations.
• 5 Small Language Models for Agentic Tool Calling — Small, open-weight language models now offer robust tool-calling for agentic AI, reducing reliance on large frontier models.
• Does RL Expand the Capability Boundary of LLM Agents? A PASS@(k,T) Analysis — RL expands LLM agent capabilities for compositional tool use, unlike its role in static reasoning.
• 😺 SubQ ships 12M tokens at 1/5 the cost — Subquadratic's new LLM, SubQ, offers a 12M-token context window at 1/5 the cost, challenging traditional Transformer limitations.
• Running local models is good now — Local LLMs have matured significantly, now enabling effective agentic coding and deep model introspection.
• Why Small Language Models (SLMs) Are Quietly Winning in 2026 — My Switch from Giant LLMs — Specialized small language models offer superior practicality, speed, and cost-efficiency over monolithic large models for most product use cases.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.