𝓢𝓽𝓪𝓷𝓵𝓮𝔂 𝓢𝓸𝓶𝓪𝓷

Cloud Architect - Semper Supra

[Oliver Prompts @oliviscusAI]

ICYMI DeepSeek just dropped an OCR model that makes most others look wasteful.

it's called deepseek-ocr.

3b parameters. 97% character-level accuracy with 10x input compression, zero fidelity loss.

most OCR systems burn 6,000+ tokens per page. this one does the same job in about 100.

→ tiny mode: ~64 tokens for simple layouts
→ small mode: ~100 tokens for standard reports
→ database mode: ~256 tokens for structured or multi-column docs
→ gundam mode: n×100 + 256 tokens for high-complexity pages

scales to 200k pages a day on a single a100. dual-stage architecture, vision encoder handles optical compression, moe decoder handles structured reconstruction.

outputs clean html, smiles, or json instead of raw pixel-to-text.

100% free. open source.

[z @zaynmcps]

GOOGLE ACABA DE MATAR LA EXCUSA DE “NO ENTIENDO ESTE REPO”

pegas cualquier repositorio y codewiki genera automáticamente:

→ documentación interactiva
→ diagramas de arquitectura
→ explicaciones del código
→ tutoriales paso a paso
→ un chatbot que entiende todo el proyecto

todo generado por gemini, siempre actualizado

te dejo la repo abajo 👇🏻

[Emil Kowalski @emilkowalski]

New skill: /pick-ui-library

Have your agent pick the right UI library for the task based on libraries I use and trust, instead of letting AI hand-roll a toast component or install an abandoned package.

Includes Sonner, cmdk, base-ui, Satori, and more.

https://t.co/jCwhthpi20

[Vaibhav Sisinty @VaibhavSisinty]

Baidu just open-sourced an OCR model that reads entire 40-page documents in one shot.

It's called Unlimited-OCR. 3 billion parameters but only 500 million active during inference. Runs 100% locally on your machine.

Why this matters: traditional OCR tools chop documents page by page. Tables that span two pages break. Reading order gets lost. Cross-page context disappears.

Unlimited-OCR processes the whole document at once. 32K context window. Text, formulas, tables, reading order all preserved across pages.

Output comes out as clean structured Markdown.

→ 93% accuracy on the standard benchmark. +6 points over the baseline.
→ Error rate stays below 0.11 even past 40 pages.
→ Multilingual out of the box.
→ 2.12 million downloads on Hugging Face last month. 14,600 GitHub
stars.

For context: Amazon Textract, Google Cloud Vision, and Azure Document Intelligence all charge per page. This runs locally for free.

[Tom Dörr @tom_doerr]

Builds a living knowledge graph of your multi-repo architecture directly from your actual source code and design documents.

https://t.co/0dOYZ6s6A1

[Teksart @TeksCreate]

Your AI agent is burning tokens on noise. Headroom fixes that.

It's a context compression layer that sits between your agent and the LLM — compressing tool outputs, logs, RAG chunks, and conversation history before they reach the model. Same answers, 60-95% fewer tokens on JSON data, 15-20% on code.

How it works:
- SmartCrusher for JSON → up to 95% reduction
- CodeCompressor (AST-based) for code
- Kompress-v2-base for prose
- CacheAligner stabilizes prefixes so provider KV caches actually hit
- Reversible compression (CCR) — originals cached locally, LLM retrieves on demand

Works with Claude Code, Codex, Cursor, Copilot, Aider, Gemini, Grok — wrap any agent in one command, undo with `headroom unwrap`.

Local-first. Your data never leaves. Cross-agent memory that mines failed sessions and writes corrections.

Live demo: 10,144 tokens → 1,260 tokens. Same FATAL error found.

https://t.co/YGPpJ6zFzr

[Marco Ramilli @Marco_Ramilli]

🤖 tokenslim
⭐ 17 stars

Stop wasting money on bloated LLM prompts. This high-performance Rust engine compresses your input tokens by 50–95% before they hit the API.

🔗 https://t.co/DVHiSZB4mc

#AI #MachineLearning

[Md Ismail Šojal 🕷️ @0x0SojalSec]

Crazy, 99% accurate OCR reasoning model like pro human,🔥

This 8B reasoning OCR model extracts text from images instantly with 99% accuracy.

- near-perfect OCR, Complex tables, Weird layouts
- instant extraction,
- From images to messy flawless perfect Markdown in seconds.
- Handles complex docs better than bigger models.
- and it reasons through tricky layouts like a human (but faster).
- Turns chaos into clean, usable Markdown.
- beats GPT-4o on document chaos.

[Ram Maheshwari @rammcodes]

Bruh... this open-source AI design skill makes your AI-generated UIs look far less generic. 😍

Hallmark is a design skill for Claude Code, Cursor & Codex that helps AI generate interfaces that feel intentionally designed instead of looking like typical AI outputs.

It can:
→ build beautiful new UIs
→ audit existing designs
→ redesign pages while preserving your content
→ study a URL or screenshot and extract its design DNA

If you build apps with AI, this is definitely worth trying :)

Follow me for more amazing AI, Coding & Web Dev insights 💎

[𝓢𝓽𝓪𝓷𝓵𝓮𝔂 𝓢𝓸𝓶𝓪𝓷 @micgoogle]

@jontec8 Add a conversion to docx pls. This would make sense with the true table and other graphics support

[Juan @7uanF]

Code Review Graph arma un grafo local de tu código con Tree-sitter, le da contexto preciso a Cursor, Claude Code o Copilot vía MCP. En reviews grandes la IA lee solo lo que importa en vez de escanear todo el repo. Un solo comando instala y configura.

[Orhun Parmaksız 👾 @orhundev]

Wow, I found a smart way to bookmark code! 🤯

🔖 codemark — Structural bookmarks that survive refactors, renames & reformatting.

🌳 Capture the semantic structure of what you marked via tree-sitter

🐀 Includes a TUI for exploring the codebase!

🦀 Written in Rust & built with @ratatui_rs

⭐ GitHub: https://t.co/c3S6VP16JX

#rustlang #ratatui #tui #coding #bookmark #semantic #devtools #programming

[Superman @thesupermanmx]

China open-sourced a peanut-sized OCR that parses entire 100-page PDFs in one shot..

It's called Unlimited-OCR. Only 3B params. Runs locally.

Every other OCR tool chops your doc into pages and loses the thread. this one reads the whole thing in a single pass.

→ One-shot "long-horizon" parsing (32K context window)
→ Multilingual, out of the box
→ 93% on the standard parsing benchmark (+6 over baseline)
→ <0.11 error rate past 40 pages
→ Runs 100% locally on your own hardware
→ Works with Transformers, vLLM, SGLang, Docker, Ollama, llama.cpp

Traditional cloud OCR (Textract, Google Vision, Azure Doc Intelligence) costs $1.50–$15 per 1,000 pages.

This runs on your machine. For free. Forever.

Baidu built it explicitly to push DeepSeek-OCR one step further. Already at 1.9M downloads on Hugging Face and most people have no idea it exists yet.

100% open source.

[Erick @ErickSky]

Ahorra tus valiosos tokens con esta herramienta que construye un grafo local de tu codebase para que tus agentes lean solo lo que realmente importa.

Resultados reales:
→ 82x menos tokens (mediana)
→ Hasta 528x en el mejor caso
→ Updates en <2 segundos
→ Todo local (cero código sale de tu máquina)
→ 30 tools MCP listas

REPOOO👇

[Arham Amin @arhamamiin]

Day 4 of learning AI bs from scratch

> understanding forward propagation and backward propagation
> got an overview of LLM’s (GPT’s progression from simple Neural networks)
> some @karpathy lectures (found a really good talk he gave for Microsoft, link below)
> layers of a modern LLM (from data collection to reward modeling to reinforcement learning)
> tried micro grad by karpathy
> image classification Neutral Network from yesterday got 92% accuracy

I think moving forward without understanding the math here intuitively is pointless, so going to take a deep dive into math around forward prop and backward propagation, major lock in needed next.

Also going to be storing all the material am consuming in a blog post and share that here.

[K2S @k2sbhai]

Moonshot AI is casually giving developers free daily access to Kimi K3 😳

no subscription
no upfront payment
just sign in and start using one of the largest open AI models available

what you get for $0:

- Kimi K3 with 2.8T parameters
- 1M token context window
- strong coding and reasoning performance
- native vision capabilities
- free daily credits that refresh automatically

why this is worth checking:

> access a frontier model without paying API fees
> long context for large codebases and documents
> works on web, desktop, mobile, and CLI

getting started takes less than 2 minutes:

1. go to https://t.co/xSpChivQtj
2. create a free account
3. Kimi K3 is available as the default model
4. start chatting or coding with your daily free credits

bonus:

Moonshot Together lets you invite friends for a chance to earn 3, 7, 15, 30, or even 365 days of Kimi Membership through its rewards program

benchmark highlights:

> 2.8T parameter MoE model
> 1M context window
> strong performance across coding, browsing, and reasoning benchmarks

important:

free credits reset daily, rate limits apply on the free tier, and the open-weight release is expected on July 27

A simple way to try one of the latest frontier AI models without paying for API access

[Ritik ☄️ @heyyritik_]

someone built an MCP that lets Claude Code rip the entire design system off any website. colors, typography, components, all of it.

clean UIs without staring at inspect element for 3 hours 💀

save this

[𝓢𝓽𝓪𝓷𝓵𝓮𝔂 𝓢𝓸𝓶𝓪𝓷 @micgoogle]

@tom_doerr There is also Structurizr C4

[𝓢𝓽𝓪𝓷𝓵𝓮𝔂 𝓢𝓸𝓶𝓪𝓷 @micgoogle]

@shugarDadddy @jakubkrehel share a screenshot before and after

[Geek @geekbb]

把你写的 HTML 片段直接变成能在 Word 里编辑的真实文档（不是图片），段落、表格、列表、图表都在。

https://t.co/RJ6x7YVu8O