MiniCPM-V Demo for HarmonyOS NEXT — v2.0

Major release. Brings the HarmonyOS port in lock-step with the new Android v2.0 / iOS v2.0 codebase, and folds in the never-released v1.3 video-understanding work as a bonus.

What's new (vs the public v1.2)

🧩 Master mtmd bridge (MBMtmd)

• Native layer is rebuilt on top of llama.cpp master's mtmd (multimodal) implementation. Same code path now drives iOS, Android, and HarmonyOS — the three platforms stay in lock-step from this version onward.
• ubatch is now tier-aware: split sizes adapt to device class so low-end SoCs no longer choke during prefill.

🌐 Racing dual-source download

• The in-app Model Manager races HuggingFace and ModelScope in parallel and uses whichever responds first. Massive win on networks where one of the two CDNs is slow or unreachable; cancels the loser to avoid wasting bandwidth.

🎞 Video understanding (carried over from the unreleased v1.3)

• Sample up to 64 frames uniformly from a video and feed them to the model.
• For MiniCPM-V 4.6 the context window is widened to nCtx = 8192 to fit longer multi-frame prompts.
• During visual prefill image_max_slice_nums is forced to 1 so each frame is one slice.

🇨🇳 No more "English-by-default" answers

• The default system prompt has been dropped on MiniCPM-V 4.6, both in TS/ArkTS and in the native processUserPrompt path. Chinese questions now answer in Chinese instead of being silently steered to English.

Requirements (unchanged)

• HarmonyOS NEXT, API 12 or later (e.g. nova 14 Vitality / Mate 60 / Pura 70)
• 64-bit ARM (arm64-v8a)
• ≥ 6 GB RAM recommended (≥ 8 GB for MiniCPM-V 2.6)

Install

hdc install MiniCPM-V-demo-HarmonyOS-v2.0.hap

Then open the app → Model Manager → Download (HF + MS race kicks in automatically).

Build info

• versionName = 2.0, versionCode = 6
• bundleName: com.openbmb.minicpmv

Verify

• MD5: 9aff126a47b651bce1d33890a289f319
• Size: ~39 MB

See DOWNLOAD.md for the full install guide.

MiniCPM-V Demo for Android — v2.0

Major release. The native multimodal pipeline has been re-platformed onto llama.cpp's master branch and the model download flow is now multi-source. Recommended upgrade for everyone on v1.7 / v1.9.

What's new

🧩 Master mtmd bridge (MBMtmd)

• Native layer is rebuilt on top of llama.cpp master's mtmd (multimodal) implementation. Same code path now drives iOS, Android, and HarmonyOS — the three platforms stay in lock-step from this version onward.
• ubatch is now tier-aware: split sizes adapt to device class so low-end SoCs no longer choke during prefill.

🌐 Racing dual-source download

• The in-app Model Manager races HuggingFace and ModelScope in parallel and uses whichever responds first. Massive win on networks where one of the two CDNs is slow or unreachable; cancels the loser to avoid wasting bandwidth.

🇨🇳 No more "English-by-default" answers

• The default system prompt has been dropped on MiniCPM-V 4.6, both in Java/Kotlin and in the native processUserPrompt path. Chinese questions now answer in Chinese instead of being silently steered to English.

Other changes

• All carry-over fixes from v1.9 (dynamic CPU dispatch for ggml-cpu) and the v1.8 video-understanding work (uniform-sample 64 frames, V4.6 nCtx=8192, prefill slice=1) remain in place.

Requirements (unchanged)

• Android 8.0 (API 26) or later
• 64-bit ARM (arm64-v8a)
• ≥ 6 GB RAM recommended (≥ 8 GB for MiniCPM-V 2.6)

Install

1. Download MiniCPM-V-demo-Android-v2.0.apk below.
2. Allow install from unknown source if prompted.
3. Open the app → Model Manager → Download (HF + MS race kicks in automatically).

Build info

• versionName = 2.0, versionCode = 12
• ABI: arm64-v8a only

Verify

• MD5: 43576791bfe8c24ed6909f33d8e94b63
• Size: ~17 MB

See DOWNLOAD.md for the full install guide.

MiniCPM-V Demo for Android — v1.9

Major upgrade since v1.7 (the unreleased v1.8 changes are folded in here). Recommended upgrade for everyone.

What's new

Video understanding (was v1.8)

• Sample up to 64 frames uniformly from a video and feed them to the model.
• For MiniCPM-V 4.6 the context window is widened to nCtx = 8192 to fit longer multi-frame prompts.
• During visual prefill image_max_slice_nums is forced to 1 so each frame is one slice — keeps prefill cost predictable when many frames are stacked.

Dynamic CPU dispatch for ggml-cpu (v1.9, PR #8)

• The native layer now dispatches to the best available CPU backend at runtime instead of being hard-pinned at build time. Older arm64 devices that lack i8mm / bf16 / dotprod will still pick the most compatible code path; newer SoCs continue to use the fastest one.

Requirements (unchanged)

• Android 8.0 (API 26) or later
• 64-bit ARM (arm64-v8a)
• ≥ 6 GB RAM recommended (≥ 8 GB for MiniCPM-V 2.6)

Install

1. Download MiniCPM-V-demo-Android-v1.9.apk below.
2. Allow install from unknown source if prompted.
3. Open the app → Model Manager → Download.

Build info

• versionName = 1.9, versionCode = 11
• ABI: arm64-v8a only

Verify

• MD5: b83f105d9ecce57f48f2afca2b5cdb69
• Size: ~17 MB

See DOWNLOAD.md for the full install guide.

MiniCPM-V Demo for HarmonyOS NEXT — v1.2

Bug-fix release on top of v1.1. Recommended upgrade for everyone running v1.1 — especially users that pick more than one image slice.

What's fixed

• No more ANR on multi-slice images. prefillImage is now executed on a libuv worker thread instead of the JS thread. This fixes the THREAD_BLOCK_3S ANR that fired when image_max_slice_nums > 1 — the heavy ViT prefill no longer blocks UI rendering.

Requirements (unchanged)

• HarmonyOS NEXT, API 12 or later (e.g. nova 14 Vitality / Mate 60 / Pura 70)
• 64-bit ARM (arm64-v8a)
• ≥ 6 GB RAM recommended (≥ 8 GB for MiniCPM-V 2.6)

Install

hdc install MiniCPM-V-demo-HarmonyOS-v1.2.hap

Then open the app → Model Manager → Download.

Build info

• versionName = 1.2, versionCode = 3
• bundleName: com.openbmb.minicpmv

Verify

• MD5: 390408e7cfe217a11ffb0b359b9a0807
• Size: ~29 MB

See DOWNLOAD.md for the full install guide.

MiniCPM-V Demo for HarmonyOS NEXT — v1.1

HarmonyOS NEXT (ArkTS + Native) build of the on-device MiniCPM-V demo. Shares the exact same llama.cpp core and model catalogue as the iOS / Android demos.

Highlights

• Multimodal chat with MiniCPM-V 2.6 / 4.0 / 4.6 (image + text)
• In-app Model Manager to download / sideload GGUF weights
• Same OBS direct-link URLs and MD5 hashes as the iOS / Android demos

Requirements

• HarmonyOS NEXT, API 12 or later (e.g. nova 14 Vitality / Mate 60 / Pura 70)
• 64-bit ARM (arm64-v8a)
• ≥ 6 GB RAM recommended (≥ 8 GB for MiniCPM-V 2.6)

Install

hdc install MiniCPM-V-demo-HarmonyOS-v1.1.hap

Then open the app → Model Manager → Download.

Build info

• versionName = 1.1, versionCode = 2
• bundleName: com.openbmb.minicpmv

Verify

• MD5: f3824e4b85940fcd8fbfe8f9a566519b
• Size: ~29 MB

See DOWNLOAD.md for the full install guide.

MiniCPM-V Demo for Android — v1.7

First public Android release. The app runs MiniCPM-V (2.6 / 4.0 / 4.6) fully on-device via llama.cpp — no remote server, no data leaves the phone.

Highlights

• Multimodal chat with MiniCPM-V 2.6 / 4.0 / 4.6 (image + text)
• In-app Model Manager to download / sideload GGUF weights
• Markdown rendering for streaming responses

Requirements

• Android 8.0 (API 26) or later
• 64-bit ARM (arm64-v8a)
• ≥ 6 GB RAM recommended (≥ 8 GB for MiniCPM-V 2.6)

Install

1. Download MiniCPM-V-demo-Android-v1.7.apk below.
2. Allow install from unknown source if prompted.
3. Open the app → Model Manager → Download.

Build info

• versionName = 1.7, versionCode = 9
• ABI: arm64-v8a only

Verify

• MD5: 9d16d89205cc6c43b68886a42b560600
• Size: ~16 MB

See DOWNLOAD.md for the full install guide.