Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.
Analysis5
github.com2026-07-07authorityabout the MiniCPM familyHN 2 pts · 0 comments719 words
MiniCPM Desk Pet runs the MiniCPM5-1B-GGUF model locally on macOS Apple Silicon or Windows x64 systems requiring about 2 GB disk space. The application is distributed under the GNU AGPL-3.0-only license and supports downloading models from Hugging Face or ModelScope during the initial setup. It provides agent-aware reactions to coding tools and includes a neko-style persona adapter for desktop interaction.written, reviewed, cited
- The app scans the machine for installed coding agents and prompts connection in one click.
- Response speed depends on the user's chip, memory pressure, and the selected model.
- Coding-agent reactions depend on each tool's integration behavior and may vary by version.
arxiv.org2024-04-10authorityabout the MiniCPM familyHN 2 pts · 0 comments449 words
MiniCPM introduces small language models with 1.2B and 2.4B parameters that match the performance of larger 7B to 13B LLMs. The approach employs extensive model wind tunnel experiments for stable scaling and utilizes a Warmup-Stable-Decay (WSD) learning rate scheduler to derive a higher compute optimal data-model ratio than Chinchilla Optimal. The family also includes MiniCPM-DPO, MiniCPM-MoE, and MiniCPM-128K variants.written, reviewed, cited
firethering.com2026-05-13about the MiniCPM familyHN 2 pts · 0 comments1,203 words
MiniCPM-V 4.6 is a 1.3B parameter multimodal model built to run on phones across iOS, Android, and HarmonyOS while handling image understanding, video analysis, OCR, and multi-image reasoning workloads. It introduces mixed 4x and 16x visual token compression which reduces visual encoding computation by more than 50% compared to standard approaches, enabling the full model to run on 4GB GPU memory or the GGUF variant on a 2GB CPU footprint. The model scores 13 on the Artificial Analysis Intelligence Index, outperforming Qwen3.5-0.8B on benchmarks like OmniDocBench where it reaches 84.6 against the rival's 47.0.written, reviewed, cited
- The Thinking variant pushes MathVista to 75.6 but introduces latency that reduces lightweight responsiveness.
- HallusionBench results show a hallucination rate of 30.6% compared to Qwen3.5-0.8B's 41.7%.
- The GGUF variant runs on CPU at 2GB, while quantized variants like BNB and AWQ sit at 3GB GPU memory.
github.com2024-05-29authorityabout the MiniCPM familyHN 3 pts · 1 comments1,339 words
github.com2025-01-14authorityHN 3 pts · 0 comments1,339 words
Community2
github.com2024-06-03authorityabout the MiniCPM familyHN 30 pts · 7 comments1,234 words
A GitHub issue alleges that the llama3-V project copied code and model structures from MiniCPM-Llama3-V 2.5, noting that weight delta means are around 1e-4 and standard deviations cluster between 1.2e-3 and 1.4e-3. The accuser states the llama3-V implementation uses a Perceiver resampler with a single-layer cross-attention instead of a two-layer self-attention and incorrectly applies Sigmoid activation of SigLIP for training multimodal large language models. The MiniCPM-Llama3-V 2.5 team confirmed they are investigating the situation.written, reviewed, cited
- The llama3-V tokenizer file size differs completely from the pre-release MinicPM-V-2 tokenizer.
- The llama3-V project claims to reference LLaVA-UHD while omitting its identical implementation to MiniCPM-Llama3-V 2.5.
old.reddit.com2024-02-02about the MiniCPM familyHN 2 pts · 2 comments
