Josh retweeted
Turned an ESP32 into a Bluetooth "finger" for Android 📱
PC (Python) → USB serial → ESP32 → BLE HID touchscreen + keyboard
Tap, swipe, pinch-zoom, Home/Back, and typing. Nothing to install on the phone and no ADB, just pair once.
I built and tested it myself on a Pixel .
just for fun ~
#ESP32 #BLE #Android #Python
Josh retweeted
It started as a weekend project to learn local inference. Now, it's a growing community.
Thank you for 1,000 GitHub stars and 100+ forks!
A special thanks to all the contributors and everyone who spent time testing and sharing ideas.
This is just the beginning!
Long-context training: don't stop too early 📈
New empirical scaling law on long-context performance:
• Scaing Law on log(LongPPL) ≈ f(N, Ds, Dl): additive power law, extrapolates to held-out scales (R²=0.97)
• Constant-Last4 protocol: stable monitoring on one extensible training trajectory
• Gains transfer: 32K → 128K, after SFT, and in long-context and agentic tasks
📄 arxiv.org/abs/2610.29352
💻 github.com/thunlp/dont-stop-…
🤗 huggingface.co/collections/E…
#LLM #LongContext #ScalingLaw #ContinualPretraining #AI #Agentic
This is awesome, Josh! 🚀 Love how MiniCPM cook makes it easier for AI agents to navigate the MiniCPM family and fine-tuning workflows. Thanks for building this and sharing it with the community! 💙
Want your agent to cook with @OpenBMB MiniCPM models?
I've compiled a dsh-plugin that helps you fine tune & work with the whole MiniCPM family model.
Open to contributions and feedback - especially to make this adapted to other coding harnesses.
This will save your agent a lot of work by having it a plugin instead of searching externally to find model data.
github.com/RegeneratusLabs/d…
Josh retweeted
ERNIE 5.5 Flash Preview scored 1668 on the latest Code Arena WebDev leaderboard, ranking #2 among open-source models and #4 among Chinese models.
This is a good step forward, and we’ll keep working to improve the model’s coding capabilities.
Ernie 5.5 Flash Preview by @Baidu_Inc is trending towards becoming the #2 open model on the Code Arena: WebDev, at 1668 pts. The weights will be open-sourced under Apache 2.0 soon.
Its score puts it just above:
- DeepSeek-V4.1-Flash (Max) at 1664
- Kimi K3 (Max) at 1658
and on par with:
- Qwen3.8-Max at 1670
Congrats to the @Baidu_Inc team on this release!
Small model showdown incoming.
@Krafton_AI Raon-1 nano launch chart puts it ahead of Qwen3.5 2B, Gemma 4 E2B, @OpenBMB MiniCPM5 2B and @liquidai LFM3 1.7B on instruction following, and it's ~2GB at Q4.
Lab numbers are a good start, but let's see how it holds up on a consumer GPU with a personal agent workflow.
All five go through the same tests on my RTX 3080, with identical llama.cpp settings:
- PA (Personal Agent) Core agent tasks
- Tool calling discipline
- Context retrieval
- Speed and VRAM fit
Which one are you betting on?👇
Is this the newest and best 2B model available?
The company behind PUBG just turned a 30B MoE into a small AI model you can download. 👀
KRAFTON released Raon-1 nano, designed for conversations, instruction following, and tool use in English and Korean.
They pruned and distilled Raon-1 mini preview, converted it into a dense model, then continued training on 3 trillion tokens before further post-training.
A smaller assistant that follows instructions and keeps track of the conversation could be useful without needing a giant model behind every interaction.
KRAFTON’s reported results ...
📋 IFBench → 75.9
🧠 PersonaMem v2, 32K context → 38.8
📱 AndroidControl → 36.0
It leads the selected comparison models on those tests—not every benchmark.
Release stats ...
💾 Downloadable weights
🔓 Apache 2.0
🌐 English + Korean
🧠 3.1B total parameters, described as an effective 1.5B dense model with per-layer embeddings
GGUFs only <4GB
So now we have to test how fast it runs, how much memory does it need, and how reliably does it handle your tools?
Josh retweeted
I just created a PlaudNote clone using a @waveshare00's ESP32-S3 based device
I was thinking of buying a device to record some meetings and take notes using voice, a dedicated device to gather some context, but the prices and subscription costs were not worth it for me, so I decided to make my own.
It is simple: you can record voice using the two internal mics, and on the server side, it transcribes the audio file using Whisper, creates a summary using a local LLM—in this demo I am using @OpenBMB MiniCPM-1B—and syncs the result with my Obsidian vault. Everything is processed locally, with no cloud and no internet.
Contributors and makers: I'll be open-sourcing the repo soon.
Josh retweeted
Introducing the MiniCPM-OCR plugin for Hermes Agent.
This is a problem I have experienced myself: LLMs by themselves are not always very good at OCR; sometimes I have needed another model specialized in it to solve complex document parsing.
For example, recently I have had difficulties understanding some physician recipes and other manuscripts.
This is where MiniCPM-OCR comes in. You can use the model and harness you already have to utilize a best-in-class OCR model by @OpenBMB.
@Teknium @NousResearch
Code: github.com/PovedaAqui/minicp…
PR: github.com/NousResearch/herm…
Josh retweeted
Introducing ERNIE Evolve — our automated research system for recursive self-improvement (RSI).
Across the three metrics reported for NanoChat AutoResearch and NanoGPT Speedrun, ERNIE Evolve surpassed the best publicly verified results available at the time of evaluation — and it already helped build ERNIE 5.5 Flash.
Excited to see MiniCPM-OCR integrated into Docling! 🚀💙 Thanks for making open-source document processing even more accessible! 🙌
Docling v2.149.0 just shipped and who should update depends on what you feed it 📄
Building RAG over PDFs?
New opt-in --use-embedded-picture-images pulls a figure straight out of the PDF at native pixels instead of a rendered crop. If the region also holds text or vector shapes it falls back to the crop
Need OCR on one consumer GPU?
New minicpm_ocr preset for OpenBMB's 1.3B MiniCPM-OCR (English + Chinese):
docling doc.pdf --pipeline vlm --vlm-model minicpm_ocr
Transcribing audio on a Mac?
ASR can now run on mlx-audio with no torch, and Granite Speech recordings get split into ~30s windows
⚠️ Scripts: --vlm-model X now really switches to the VLM pipeline (it used to be silently ignored), and pairing it with --pipeline standard now errors
pip install -U docling
github.com/docling-project/d…
Josh retweeted
Holy crap, a fully localized, privacy-first OCR small model stack actually got polished into a usable personal product by an independent dev!
Josh (@RightSideOfAI)just released v1.0 powered by MiniCPM5-2B and MiniCPM-OCR. Simply put, it combines an on-device lightweight LLM with a specialized text recognition OCR, running entirely on the user's local hardware without hitting the cloud.
In an era dominated by cloud APIs and constant data privacy concerns, taking the "purely local, lightweight tool" approach is a smart move. When processing sensitive documents, receipts, or private notes, users' biggest hesitation has always been sending data to remote servers. This MiniCPM combo hits a sweet spot between compact footprint, local runnable performance, and high recognition accuracy.
That said, for a pure local tool, launching v1.0 is just the first step. What really determines user retention down the road will be its power efficiency on lower-spec machines, accuracy with complex layouts or long-form documents, and actual response speed.
If future updates keep resource usage lean, this kind of lightweight, zero-privacy-risk on-device AI tool could easily become a must-have setup for devs and privacy-conscious users alike.
@taylor_ @ChrisUniverse
materia release 1.0.0: blood panel in, cited supplement and herb options out.
every claim traceable to a source, withheld when the citation doesn't hold.
100% open weights on CPU: @OpenBMB MiniCPM5-2B for chat and reranking, MiniCPM-OCR-1.0 for lab-report OCR, llama.cpp, SQLite BM25, a deterministic safety screen.
Not a diagnosis, not a prescription. Local, no cloud, 100% private.
github.com/RegeneratusLabs/m…
Two awesome builds! 🔥 Love seeing MiniCPM power such different use cases, all running locally and offline. Thanks for building and sharing! 💙🚀
My @OpenBMB MiniCPM Spark S2 program submissions! 2 different angles - help the health of a human or help the memory of an agent.
all offline, all local, all powered & optimised for the MiniCPM family!
Lets gooo!
Get your submissions in.
github.com/RegeneratusLabs/m…
blood panel in, cited supplement and herb options out. a naturopathic or integrative doctor AI assistant that works completely offline.
every claim traceable to a source, withheld when the citation doesn't hold.
100% open weights on CPU: OpenBMB MiniCPM5-2B for chat and reranking, MiniCPM-OCR-1.0 for lab-report OCR, llama.cpp, SQLite BM25, a deterministic safety screen.
github.com/RegeneratusLabs/m…
Your agent forgets the day; this gives it back. A nightly pass reads the day's conversations, keeps the facts that matter, and writes a diary you can read — all on your own machine with a small MiniCPM model.
Love this idea! 💙 Giving agents a way to organize their memories offline with MiniCPM is such a creative use case. Excited to see minicpm-dream in Spark S2! 🚀
Next @OpenBMB based project
minicpm-dream.
offline dream memory tool. your agent forgets the day; this gives it back. A nightly pass (based on the existing dreaming memory protocol) it reads the day's conversations, keeps the facts that matter, and writes a diary you can read all on your own machine with a small MiniCPM model.
currently optimised to plug into @NousResearch Hermes - but open to contributions for more harnesses.
offline local models sorting your memories while you sleep.
I will be entering this into the OpenBMB Sprark Program S2 hackathon! Alongside my recent materia project.
follow along for more fun small language model projects
github.com/RegeneratusLabs/m…
Awesome work, Josh! 💙 Love seeing MiniCPM5-2B and MiniCPM-OCR come together in a fully local, privacy-focused application. Congrats on the v1.0 release! 🚀
materia release 1.0.0: blood panel in, cited supplement and herb options out.
every claim traceable to a source, withheld when the citation doesn't hold.
100% open weights on CPU: @OpenBMB MiniCPM5-2B for chat and reranking, MiniCPM-OCR-1.0 for lab-report OCR, llama.cpp, SQLite BM25, a deterministic safety screen.
Not a diagnosis, not a prescription. Local, no cloud, 100% private.
github.com/RegeneratusLabs/m…
Josh retweeted
Replying to @Teknium
Shadow-git checkpoints before every modifying tool call → autonomous rollback when the agent wrecks the workspace → a timeline pane to scrub and restore history.
teapawt.net/MossAIStudio/status/21…
I built Hermes Chronos for the @NousResearch Hermes Agent Plugin Hackathon.
Shadow-git checkpoints before every modifying tool call → autonomous rollback when the agent wrecks the workspace → a timeline pane to scrub and restore history.
88s demo. Sound on. 🎧
#HermesAgent
Josh retweeted
We're giving away a Hogwarts Legacy custom wrapped NVIDIA GeForce RTX 5090 to celebrate its official release with DLSS 5 🪄
To enter:
🟢 Share this post
🟢 Comment #RTXON
Terms and Conditions: nvidia.com/en-us/geforce/con…
Josh retweeted
ERNIE 5.5 Flash Base just dropped! Time to open that project folder. 🚀
Our efficient sparse MoE language model draws on data spanning multiple domains and languages, achieving SOTA on multiple benchmarks among similarly sized base models. Its architecture and training system were designed together for efficient training and inference, with advances in attention mechanisms, expert routing, and training objectives:
- 280B total parameters; 14B activated per token
- Pre-trained on 33.5T tokens
- Up to 1M-token context for fine-tuning and inference
- FP8 and BF16 weights
The weights come straight from pre-training, with no mid- or post-training. Compare methods from the same checkpoint, measure what each adds, and shape the model around your own tasks.





































