4B Parameter Agent Model

LocoOperator
— 4B —

A distilled code exploration agent trained on Qwen3-Coder-Next traces. Navigate codebases with precision — locally, at zero cost.

Samples 170K
JSON Valid 100%
Tool Align 100%
Syntax OK 100%
View on GitHub
Architecture
Agent Collaboration

A two-tier agent system where the main model delegates codebase exploration to a lightweight local sub agent.

◆

Qwen3-Coder-Next

Main Agent

Decision-making, planning, code generation and editing. Orchestrates the full agent loop.

◈

LocoOperator-4B

Explorer Sub Agent

Codebase search, file reading, code navigation and analysis. Fast, focused, local.

GGUF llama.cpp Mac Studio 50K ctx 10 turns
Training
Specifications
Base Model
Qwen3-4B-Instruct-2507
Teacher
Qwen3-Coder-Next
Method
Full-Parameter SFT
Training Data
170,356 samples
Hardware
4× H200 141GB
Parallelism
DDP
Precision
BF16
Epochs
1
Effective Batch
32
Learning Rate
2e-5
Max Seq Length
16,384 tokens
Training Time
~25 hours
Template
qwen3_nothinking
Framework
MS-SWIFT
Checkpoint
Step 2524
Evaluation
Results on 65 Samples
0%
Tool Call Presence Alignment
65/65
0%
First Tool Type Match
40/61
0%
JSON Validity
76/76
0%
Argument Syntax Correctness
76/76
Comparison
Tool Distribution
LocoOperator-4B
Qwen3-Coder-Next
ModelJSON ValidArgs SyntaxIssues
LocoOperator-4B76/76 (100%)76/76 (100%)None
Qwen3-Coder-Next89/89 (100%)78/89 (87.6%)11 calls with empty arguments
Caveats
Known Limitations