Modelpublic

Arm/qwen3-0-6b-base-q4-k-m-llamacpp-raspberrypi5

sourceHugging Faceapache-2.0updated 12d agoView on Hugging Face

0likes708downloads

Model Card

Qwen3-0.6B-Base optimized for Arm-based Edge Linux

Qwen3-0.6B-Base text generation optimized as a K-quant GGUF model (llama.cpp) for Arm-based Edge Linux systems.

Summary

This repository contains an Arm-optimized version of Qwen/Qwen3-0.6B-Base for text generation. The model is provided in GGUF format (llama.cpp runtime), targeting Edge Linux systems. Weights are quantized via llama.cpp K-quant (Q4KM mixed recipe: predominantly Q4K, with Q6K and Q8_0 on select tensors, F32 norms); activations are unquantized, computed in ggml's native kernels at runtime.

This version demonstrates efficient inference on Arm-based platforms while preserving the original model's intended behavior. Arm evaluated this model on MMLU and measured performance on a representative evaluation target.

Key results

Area	Result
Model format	GGUF (llama.cpp)
Target device class	Edge Linux
Reference device	Raspberry Pi 5 (Cortex-A76, Linux Raspberry Pi OS 64-bit, based on Debian 13 "Trixie")
Primary performance result	25.26 tokens/sec, 3635.67 ms TTFT
Accuracy result	MMLU 51.75% (5-shot)
Size / memory result	414.26 MB, 2.76x smaller than the bf16 baseline

Original model

Field	Value
Original model	Qwen/Qwen3-0.6B-Base
Original source	Hugging Face
Original developer	Alibaba Cloud (Qwen team)
Original model card	Qwen/Qwen3-0.6B-Base
Original license	Apache-2.0

Model files

File	Description
`Qwen__Qwen3-0.6B-Base_llamacpp_optimized.gguf`	Arm-optimized model for deployment
`example.py`	Minimal inference example
`pyproject.toml`	Python dependencies and deployment requirements for the example
`uv.lock`	Locked dependency versions for the example
`.python-version`	Pinned Python version (3.12) for the example
`config.yaml`	Model I/O contract used by the example
`benchmarks/`	Benchmark scripts and raw results

Performance

Performance was measured on the reference configuration below. Results are intended to make the optimization reproducible but do not guarantee identical performance on every Arm-based system.

Reference configuration

Field	Value
Device / platform	Raspberry Pi 5
CPU / accelerator	Cortex-A76 (4 cores @ 2.4 GHz), CPU
OS	Linux — Raspberry Pi OS 64-bit, based on Debian 13 "Trixie"
Runtime	llama.cpp
Backend / delegate	N/A
Batch size	1
Precision	weights quantized via llama.cpp K-quant (Q4KM mixed recipe: predominantly Q4K, with Q6K and Q8_0 on select tensors, F32 norms); activations unquantized
Runs	5 measured runs

Performance results

Metric	Original / baseline	Arm-optimized	Improvement
p50 latency	146581.98 ms	8702.57 ms	16.84x
TTFT	108482.43 ms	3635.67 ms	29.84x
Tokens per second	3.36	25.26	7.52x
Model size	1142.68 MB	414.26 MB	2.76x smaller
Peak memory	1276.52 MB	946.12 MB	1.35x less

Accuracy

Accuracy was evaluated using the same preprocessing, input resolution, and evaluation protocol described below. Where possible, the optimized model is compared against the original model under the same evaluation conditions.

Evaluation setup

Field	Value
Dataset	MMLU
Split	test (5-shot)
Number of samples	13508
Metric(s)	Accuracy
Evaluation runtime	llama.cpp

Accuracy results

Metric	Original / baseline	Arm-optimized	Change
MMLU accuracy (5-shot)	52.98%	51.75%	-1.23 pp

Accuracy was measured using the evaluation setup described above. Users should re-evaluate the model on their own data before production use.

Arm optimization approach

Arm optimized this model for efficient inference on Arm-based platforms using a hardware-aware conversion and validation flow.

For this release, Arm used:

Optimization area	Applied?	Notes
Model conversion	Yes	Converted from the Hugging Face checkpoint to a bf16 GGUF reference via converthftogguf.py, then quantized to GGUF Q4K_M via llama-quantize
Quantization	Yes	llama.cpp K-quant post-training weight quantization (Q4KM mixed recipe: predominantly Q4K, with Q6K and Q8_0 on select tensors, F32 norms), weight-only — activations unquantized; guided by an importance matrix (llama-imatrix) computed over a WikiText-2 calibration corpus
Runtime/backend selection	Yes	llama.cpp / ggml CPU (arm64); this Q4KM recipe's Q4K and Q6K tensors fall back to generic ggml aarch64 kernels rather than an Arm KleidiAI acceleration path
Graph/runtime compatibility updates	Yes	Performed as part of the llama.cpp GGUF conversion pipeline (converthfto_gguf.py + llama-imatrix + llama-quantize)
Accuracy validation	Yes	Compared against the original model or published baseline
Performance validation	Yes	Measured on the reference Arm platform

The goal of this process is to improve deployment characteristics such as latency, memory use, model size, and runtime compatibility while preserving the model's intended behavior. Detailed conversion scripts, calibration configuration, or backend-specific implementation details may be provided separately where appropriate.

Using this model

Requirements

—Python 3.12 (see .python-version), managed via uv.
—llama-cpp-python has no prebuilt wheel for this platform and is compiled from source by uv sync. This requires a C/C++ toolchain and CMake:
—Raspberry Pi OS 64-bit, based on Debian 13 "Trixie" (arm64): build-essential, cmake
—Ubuntu (Noble, arm64): build-essential (12.10ubuntu1), cmake (3.28.3-1build7)

Install dependencies

bash

uv python install
uv sync --frozen

Run the example

bash

uv run example.py

To save the predictions as JSON, provide an output path explicitly:

bash

uv run example.py --output predictions.local.json

Expected input

Property	Value
Input shape	[1, 2048]
Input type	int64
Input range	[0, 151935]
Preprocessing	Token ids produced by the byte-level BPE tokenizer embedded in the .gguf (tokenizer.ggml.model=gpt2, pre-tokenizer=qwen2, vocab 151936); bostokenid 151643; no chat template; max context 32768 tokens

Expected output

Property	Value
Output shape	[1, 151936]
Output type	Next-token logits over the 151936-entry vocabulary at the final position
Postprocessing	Greedy sampling (temperature 0.0), eostokenid 151643, detokenized with the same embedded byte-level BPE tokenizer

Intended use

This model is intended for developers evaluating text-generation workloads on Arm-based platforms. It is suitable as a reference implementation for benchmarking, prototyping, and integration exploration.

Limitations

—Performance depends on the target device, runtime version, backend/delegate support, memory configuration, and system load.
—Accuracy was evaluated on MMLU (test, 5-shot) and may not generalize to all domains.
—This release preserves the original model's intended task and behavior, but users should validate it for their own application, data, and deployment environment.
—This repository is not a replacement for the original model documentation.

Additional notes

—This is the base (pretrained) checkpoint, not an instruction-tuned variant — it does not reliably follow chat-style instructions and has no chat template; use few-shot prompting or an instruct checkpoint for conversational use.
—MMLU accuracy above was measured on a capped evaluation set (13508 of the full test split).
—The bf16 baseline is a reference point rather than a practical deployment option on this device class — it exists to make the quantization delta measurable on identical hardware.

About this version

Original Model: Qwen/Qwen3-0.6B-Base by Alibaba Cloud (Qwen team) - Repository

Optimization/conversion: Arm-Optimized version for execution on Arm-based platforms.

Converted/optimized by: Arm

License: The Original Model and the Optimized Model are subject to Apache-2.0.

This repository contains a converted or optimized version of the Original Model (the “Optimized Model”). The Original Model has been converted or optimized as described above for execution on Arm-based platforms.

No retraining or fine-tuning of the Original Model was performed as part of the conversion or optimization. The conversion or optimization was not intended to change the Original Model’s behavior or intended use.

Original Model and Documentation

For information about the Original Model, including its development, training data, intended uses, limitations and other relevant information, please refer to the Original Model repository. Information in that repository was provided by the original developer or other third parties and, unless expressly stated otherwise, has not been independently verified by Arm.

Licenses and Third-Party Terms

Use of the Original Model and the Optimized Model is subject to the applicable licenses, usage restrictions and other terms identified above and in the relevant repositories. Publication of the Optimized Model does not grant any rights beyond those provided under the applicable license terms.

You are responsible for reviewing those terms and ensuring that your use of the Original Model and the Optimized Model is permitted.

Purpose of this Release

The Optimized Model is provided as a reference implementation to demonstrate and evaluate execution and performance on Arm-based systems. It is not a production-ready or supported solution.

Arm’s publication of the Optimized Model does not constitute an endorsement or certification of the Original Model or a representation that the Optimized Model is suitable for production use or any particular purpose.

To the fullest extent permitted by applicable law (i) the Optimized Model is provided “as is.” Arm makes no representations or warranties that the Original Model, the Optimized Model or their outputs are accurate, safe, secure, non-infringing, legally compliant, suitable for production use or fit for any particular purpose; and (ii) Arm will not be liable for any loss or damage arising from or in connection with the Optimized Model, its use or its outputs.

You are responsible for independently evaluating the Optimized Model, its outputs and its suitability for your intended use, including compliance with applicable legal, regulatory, safety and security requirements.

Arm does not commit to provide ongoing support, maintenance or updates for the Optimized Model. Any use of or reliance on the Optimized Model or its outputs is at your own risk.