Can a 4B Small Model Replace Top‑Tier Closed‑Source LLMs? Microsoft’s Terminus‑4B Cuts Token Use by 30%

Large language models often hit context‑length limits when handling verbose terminal logs during code‑related tasks, forcing the main model to process massive amounts of raw output and reducing efficiency.

Execution Subagent with a Tiny Model

Microsoft introduces an Execution Subagent built around a 4B‑parameter model called Terminus‑4B. The sub‑agent receives simple commands (e.g., run a test suite) and operates in its own context, returning a fixed‑format summary that includes the command run, result, and key error locations, allowing the main model to work with a concise 200‑word report.

Specialized Small‑Model Training

The team collected roughly 3,200 real terminal tasks from open‑source projects, covering five major languages and focusing on test execution and error diagnosis. Training proceeded in two stages: supervised fine‑tuning (SFT) on internal telemetry data for two epochs, followed by reinforcement learning (RL) using the GRPO algorithm, which steadily raised reward scores and taught high‑value strategies.

Benchmark Results

On the SWE‑Bench Pro and internal SWE‑Bench C# benchmarks, Terminus‑4B as the sub‑agent paired with Claude Opus 4.6 as the main model achieved a system‑wide success rate of 31.5% (baseline 30.0%) and reduced overall token usage by roughly 13%, a net 30% cut compared with using Opus directly.

The sub‑agent also lowered the main model’s direct terminal calls by 73.7%. Further experiments with different main models (Opus, GPT‑5.3‑Codex) showed similar token savings and reduced command invocations. An extreme test removing the main model’s own terminal tool forced all work through the sub‑agent; an untrained Vanilla‑4B then caused a 9.5% token increase and a 1.51× higher sub‑agent call rate, whereas the RL‑fine‑tuned Terminus‑4B matched Opus’s best performance.

Conclusion

Delegating noisy, repetitive terminal work to a cheap, purpose‑built small model preserves the capabilities of expensive top‑tier models for high‑level reasoning, offering a practical path to lower the barrier for widespread autonomous programming assistants.