Model	Size_GB	BPW	PPL_Q	KLD_Mean	KLD_Max	Top_P_Match
LocoOperator-4B-BF16.gguf	7.498	16.01	9.24309	-1.2e-05	4e-06	100.000%
LocoOperator-4B-MXFP4_MOE.gguf	3.986	8.51	9.24606	0.001835	2.98238	97.518%
LocoOperator-4B-i1-MXFP4_MOE.gguf	3.986	8.51	9.24606	0.001835	2.98238	97.518%
LocoOperator-4B-Q8_0.gguf	3.986	8.51	9.24606	0.001835	2.98238	97.518%
LocoOperator-4B-i1-Q8_0.gguf	3.986	8.51	9.24606	0.001835	2.98238	97.518%
LocoOperator-4B-Q6_K.gguf	3.079	6.58	9.27926	0.0068	10.5686	95.526%
LocoOperator-4B-i1-Q6_K.gguf	3.079	6.58	9.295	0.006075	15.9945	95.857%
LocoOperator-4B-i1-Q5_1.gguf	2.841	6.07	9.28859	0.01364	2.98838	94.135%
LocoOperator-4B-Q5_1.gguf	2.841	6.07	9.43222	0.022675	16.3454	93.161%
LocoOperator-4B-Q5_K_M.gguf	2.691	5.75	9.35457	0.017023	12.3947	93.635%
LocoOperator-4B-i1-Q5_K_M.gguf	2.691	5.75	9.2965	0.013153	7.78613	94.257%
LocoOperator-4B-i1-Q5_0.gguf	2.636	5.63	9.42255	0.019663	17.94	93.208%
LocoOperator-4B-Q5_0.gguf	2.63	5.62	9.41521	0.023403	31.4019	92.839%
LocoOperator-4B-Q5_K_S.gguf	2.63	5.62	9.44087	0.022119	13.6483	92.800%
LocoOperator-4B-i1-Q5_K_S.gguf	2.63	5.62	9.28767	0.014865	7.65169	93.702%
LocoOperator-4B-Q4_1.gguf	2.418	5.16	9.66722	0.074718	15.0861	87.757%
LocoOperator-4B-i1-Q4_1.gguf	2.418	5.16	9.45293	0.038707	13.8444	90.574%
LocoOperator-4B-Q4_K_M.gguf	2.326	4.97	9.48239	0.048236	15.3105	90.300%
LocoOperator-4B-i1-Q4_K_M.gguf	2.326	4.97	9.48582	0.03368	13.551	91.233%
LocoOperator-4B-IQ4_NL.gguf	2.229	4.76	9.60891	0.050173	11.4324	89.708%
LocoOperator-4B-i1-Q4_K_S.gguf	2.22	4.74	9.47603	0.039843	10.0551	90.557%
LocoOperator-4B-Q4_K_S.gguf	2.22	4.74	9.80236	0.068821	15.209	88.513%
LocoOperator-4B-i1-IQ4_NL.gguf	2.218	4.74	9.50223	0.039414	8.18964	90.573%
LocoOperator-4B-i1-Q4_0.gguf	2.213	4.73	9.79026	0.063915	12.6928	88.737%
LocoOperator-4B-Q4_0.gguf	2.207	4.71	9.86629	0.074527	13.2501	87.758%
LocoOperator-4B-IQ4_XS.gguf	2.129	4.55	9.62193	0.051911	11.0682	89.705%
LocoOperator-4B-i1-IQ4_XS.gguf	2.115	4.52	9.49687	0.040098	7.03875	90.402%
LocoOperator-4B-Q3_K_L.gguf	2.086	4.45	10.2476	0.121944	27.0257	84.146%
LocoOperator-4B-i1-Q3_K_L.gguf	2.086	4.45	9.90811	0.090874	15.8122	86.154%
LocoOperator-4B-Q3_K_M.gguf	1.933	4.13	10.7021	0.15788	20.2044	82.662%
LocoOperator-4B-i1-Q3_K_M.gguf	1.933	4.13	9.98057	0.102708	16.8243	85.354%
LocoOperator-4B-i1-IQ3_M.gguf	1.828	3.9	10.1634	0.137347	14.6883	83.180%
LocoOperator-4B-IQ3_M.gguf	1.828	3.9	14.2539	0.557713	19.4397	67.631%
LocoOperator-4B-IQ3_S.gguf	1.769	3.78	15.0624	0.619131	20.122	65.931%
LocoOperator-4B-i1-IQ3_S.gguf	1.769	3.78	10.1755	0.142066	17.0028	83.139%
LocoOperator-4B-i1-Q3_K_S.gguf	1.757	3.75	10.8886	0.171224	28.3373	82.133%
LocoOperator-4B-Q3_K_S.gguf	1.757	3.75	11.5475	0.237895	30.6868	79.412%
LocoOperator-4B-i1-IQ3_XS.gguf	1.69	3.61	10.3629	0.168783	14.3358	81.928%
LocoOperator-4B-i1-Q2_K.gguf	1.555	3.32	12.1574	0.328652	18.6622	75.570%
LocoOperator-4B-i1-IQ3_XXS.gguf	1.555	3.32	11.2795	0.263448	25.251	77.569%
LocoOperator-4B-Q2_K.gguf	1.555	3.32	17.153	0.713596	16.3946	64.880%
LocoOperator-4B-i1-Q2_K_S.gguf	1.456	3.11	13.1709	0.450125	18.3826	71.231%
LocoOperator-4B-i1-IQ2_M.gguf	1.409	3.01	14.0857	0.544764	18.5618	67.933%
LocoOperator-4B-i1-IQ2_S.gguf	1.32	2.82	15.0717	0.621189	24.0981	65.722%
LocoOperator-4B-i1-IQ2_XS.gguf	1.261	2.69	16.8277	0.750336	19.2128	63.162%
LocoOperator-4B-i1-IQ2_XXS.gguf	1.161	2.48	27.5988	1.32144	14.6807	52.522%
LocoOperator-4B-i1-IQ1_M.gguf	1.05	2.24	49.0978	1.9323	16.5947	44.067%
LocoOperator-4B-i1-IQ1_S.gguf	0.983	2.1	139.951	3.03274	16.0947	28.387%

MODEL

Introduction

LocoOperator-4B is a 4B-parameter tool-calling agent model trained via knowledge distillation from Qwen3-Coder-Next inference traces. It specializes in multi-turn codebase exploration — reading files, searching code, and navigating project structures within a Claude Code-style agent loop. Designed as a local sub agent, it runs via llama.cpp at zero API cost.

	LocoOperator-4B
Base Model	Qwen3-4B-Instruct-2507
Teacher Model	Qwen3-Coder-Next
Training Method	Full-parameter SFT (distillation)
Training Data	170,356 multi-turn conversation samples
Max Sequence Length	16,384 tokens
Training Hardware	4x NVIDIA H200 141GB SXM5
Training Time	~25 hours
Framework	MS-SWIFT

Key Features

Tool-Calling Agent: Generates structured <tool_call> JSON for Read, Grep, Glob, Bash, Write, Edit, and Task (subagent delegation)
100% JSON Validity: Every tool call is valid JSON with all required arguments — outperforming the teacher model (87.6%)
Local Deployment: GGUF quantized, runs on Mac Studio via llama.cpp at zero API cost
Lightweight Explorer: 4B parameters, optimized for fast codebase search and navigation
Multi-Turn: Handles conversation depths of 3–33 messages with consistent tool-calling behavior

Performance

Evaluated on 65 multi-turn conversation samples from diverse open-source projects (scipy, fastapi, arrow, attrs, gevent, gunicorn, etc.), with labels generated by Qwen3-Coder-Next.

Core Metrics

Metric	Score
Tool Call Presence Alignment	100% (65/65)
First Tool Type Match	65.6% (40/61)
JSON Validity	100% (76/76)
Argument Syntax Correctness	100% (76/76)

The model perfectly learned when to use tools vs. when to respond with text (100% presence alignment). Tool type mismatches are between semantically similar tools (e.g. Grep vs Read) — different but often valid strategies.

Tool Distribution Comparison

JSON & Argument Syntax Correctness

Model	JSON Valid	Argument Syntax Valid
LocoOperator-4B	76/76 (100%)	76/76 (100%)
Qwen3-Coder-Next (teacher)	89/89 (100%)	78/89 (87.6%)

LocoOperator-4B achieves perfect structured output. The teacher model has 11 tool calls with missing required arguments (empty arguments: {}).

Quick Start

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "LocoreMind/LocoOperator-4B"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# prepare the messages
messages = [
    {
        "role": "system",
        "content": "You are a read-only codebase search specialist.\n\nCRITICAL CONSTRAINTS:\n1. STRICTLY READ-ONLY: You cannot create, edit, delete, move files, or run any state-changing commands. Use tools/bash ONLY for reading (e.g., ls, find, cat, grep).\n2. EFFICIENCY: Spawn multiple parallel tool calls for faster searching.\n3. OUTPUT RULES: \n   - ALWAYS use absolute file paths.\n   - STRICTLY NO EMOJIS in your response.\n   - Output your final report directly. Do not use colons before tool calls.\n\nENV: Working directory is /Users/developer/workspace/code-analyzer (macOS, zsh)."
    },
    {
        "role": "user",
        "content": "Analyze the Black codebase at `/Users/developer/workspace/code-analyzer/projects/black`.\nFind and explain:\n1. How Black discovers config files.\n2. The exact search order for config files.\n3. Supported config file formats.\n4. Where this configuration discovery logic lives in the codebase.\n\nReturn a comprehensive answer with relevant code snippets and absolute file paths."
    }
]

# prepare the model input
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# conduct text completion
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512,
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()

content = tokenizer.decode(output_ids, skip_special_tokens=True)
print(content)

Local Deployment

For GGUF quantized deployment with llama.cpp, hybrid proxy routing, and batch analysis pipelines, refer to our GitHub repository.

Training Details

Parameter	Value
Base model	Qwen3-4B-Instruct-2507
Teacher model	Qwen3-Coder-Next
Method	Full-parameter SFT
Training data	170,356 samples
Hardware	4x NVIDIA H200 141GB SXM5
Parallelism	DDP (no DeepSpeed)
Precision	BF16
Epochs	1
Batch size	2/GPU, gradient accumulation 4 (effective batch 32)
Learning rate	2e-5, warmup ratio 0.03
Max sequence length	16,384 tokens
Template	qwen3_nothinking
Framework	MS-SWIFT
Training time	~25 hours
Checkpoint	Step 2524

Known Limitations

First-tool-type match is 65.6% — the model sometimes picks a different (but not necessarily wrong) tool than the teacher
Tends to under-generate parallel tool calls compared to the teacher (76 vs 89 total calls across 65 samples)
Preference for Bash over Read may indicate the model defaults to shell commands where file reads would be more appropriate
Evaluated on 65 samples only; larger-scale evaluation needed

License

MIT

Acknowledgments

Qwen Team for the Qwen3-4B-Instruct-2507 base model
MS-SWIFT for the training framework
llama.cpp for efficient local inference
Anthropic for the Claude Code agent loop design that inspired this work

SimplySara/LocoOperator-4B-GGUF

This is a static quantization of LocoreMind/LocoOperator-4B, made by SimplySara