mobilecore-qwen3-0.6b-gguf

Qwen3-0.6B GGUF Q4_K_M for MobileCore

This repository contains a Harzva-produced GGUF conversion of Qwen/Qwen3-0.6B for mobile llama.cpp runtimes.

This model is maintained as part of the Mobile Model Playground, which records conversion provenance, checksums, compatibility, and mobile validation evidence.

Status: conversion candidate · Android emulator contract checked. No physical-phone performance claim is made yet.

Distribution

• Hugging Face — international discovery and direct HTTPS download.
• GitCode — domestic distribution through Git LFS.

Both endpoints publish the same Harzva-converted artifact. Verify the byte count and SHA-256 below after downloading from either endpoint.

Artifact

Provenance

• Upstream: Qwen/Qwen3-0.6B
• Immutable upstream revision: c1899de289a04d12100db370d81485cdf75e47ca
• Source model.safetensors: 1,503,300,328 bytes
• Source SHA-256: f47f71177f32bcd101b7573ec9171e6a57f4f4d31148d38e382306f42996874b
• Converter/runtime: ggml-org/llama.cpp
• Converter revision: fd1a05791da1c06dc8bebdd537aa5212831d7cf6
• Intermediate F16 GGUF SHA-256: d04bceb664d484eaf134cdbc63745f5241bea80132c458e61c9449f488fe2abc
• Artifact origin: converted_by_harzva
• License review: Apache-2.0 redistribution cleared

Capability matrix

Validation

Host smoke

• Environment class: macOS arm64 host, Metal backend
• Prompt: Q: What is 2+2? A:
• Response: 2 + 2 = 4.
• Prompt processing: 952.8 tok/s
• Generation: 112.2 tok/s

This only verifies that the converted artifact loads and generates a sane answer on the host.

Android ARM64 emulator

• Model load: 1,389 ms
• First token: 8,026 ms
• Generation: 0.289 tok/s
• Artifact file size reported by the test: 456 MB

These numbers came from a low-performance Android ARM64 AVD. They are contract evidence, not phone performance. Peak PSS, temperature, power, sustained stability, and physical-device quality remain unverified.

Sanitized machine-readable records for the host, Android AVD, and GitCode post-publication checks are available under evidence/runs/. Hugging Face publication verification is tracked independently by the Mobile Model Playground so one distribution endpoint cannot inherit another endpoint's evidence.

Verify

shasum -a 256 -c SHA256SUMS

Download from Hugging Face with:

hf download harzva/mobilecore-qwen3-0.6b-gguf qwen3-0.6b-q4_k_m.gguf --local-dir .

Reproduce

See conversion/convert.sh. Reproduction requires Python conversion dependencies and a llama.cpp build containing llama-quantize.

MobileCore use

Import the GGUF as a text model and select the llama.cpp backend. Do not attach an mmproj or advertise visual/audio input for this artifact.

Known limitations

• Q4_K_M was produced without an importance matrix.
• Physical Android device results are not yet available.
• The surviving conversion artifacts were verified, but a fresh byte-for-byte conversion transcript is still desirable for long-term provenance.