QClaw-4B

QCLAW

QClaw-4B is a 4-billion parameter language model fine-tuned for agentic tasks and tool use, designed for use with OpenClaw-compatible agent frameworks.

Despite its compact size, QClaw-4B achieves state-of-the-art results in the 4B class, matching or exceeding models several times larger — including Kimi K2.5 and GLM-4.5 — on the ClawBench agent benchmark.

Benchmark Results

ClawBench

Metric	Value
Overall Score	84.8 / 100
Pass Rate	73.5%
Tasks Evaluated	1110
Errors / Timeouts	0
Avg. Tokens per Task	3301

Leaderboard Comparison

Model	Score	Size
Kimi K2.5	85.0	large
GLM-4.5	85.0	large
QClaw-4B	84.8	4B
GLM-4.5 Air	84.0	large

Ru-Arena-General

Model Name	Winrate	95% CI	Average # Tokens
Vistral-24B-Instruct	96.1	(-0.7, 0.8)	647
Mistral-Small-3.2-24B-Instruct-2506	92.1	(-0.9, 1.0)	486
QClaw-4B	80.7	(-1.4, 1.4)	3301
vikhr-nemo-12b-instruct-r-21-09-24(180 leaked)	79.8	(-2.2, 1.9)	627

QClaw-4B is SOTA in the ≤4B parameter class, outperforming Qwen 9B and other models more than twice its size.

Benchmark API tokens and evaluation cards provided by Aleksandr Nikolich — Love. Death. Transformers..

Model Details

Architecture: Decoder-only transformer
Parameters: ~4B
Benchmark suite: ClawBench
Primary use case: Agentic workflows, tool calling, multi-step reasoning

Training

QClaw-4B was trained on a curated mixture of:

Agentic task trajectories (tool calling, multi-step planning)
Instruction-following data
Code and structured reasoning

Training annotation cards and dataset curation provided by Aleksandr Nikolich — Love. Death. Transformers..

Intended Use

QClaw-4B is intended for:

Agentic pipelines using OpenClaw or compatible frameworks
Tool-augmented assistants requiring compact, efficient inference
Research into small-model agent capabilities

Out-of-scope use: QClaw-4B is not intended for safety-critical systems without additional alignment work.

Citation

@misc{qclaw4b2026,
  title  = {QClaw-4B: State-of-the-Art 4B Agent Model for OpenClaw},
  author = {Nikolay Kompanets (LakoMoor)},
  year   = {2026},
  url    = {https://huggingface.co/LakoMoor/QClaw-4B}
}

License

Apache 2.0

LakoMoor/QClaw-4B