19 y/o building real projects while learning web, AI, and game development. Can’t get enough of AI 😁
India
Joined April 2025
- Tweets957
- Following48
- Followers40
- Likes2.6K
Harsh 💢 retweeted
Indra OCR 1.1 is out now!
- 1M param main OCR model, 44 pages/s on CPU
- Within 4 points of 3.4B param DeepSeek OCR 2 on OmniParseBench
- 15x faster than Tesseract, and more accurate
- 80+ languages, plus handwriting
Harsh 💢 retweeted
Today we are releasing JST-2, the world's first full-stack physics-native world foundation model.
Many world models learn in pixel/video space. We went back to first principles and built the entire model stack in 3D.
Generate a 3D scene in seconds and expand it in real time. Create characters with motion built in, all through one general architecture.
Harsh 💢 retweeted
Today, we’re releasing Bonsai 2 2B, 4B, and 9B.
Based on Qwen3.5, these new ternary models bring stronger reasoning, visual understanding, and tool calling to compact footprints designed for everyday devices:
• 2B: 0.43GB
• 4B: 0.95GB
• 9B: 2.0GB
For this release, we focused on what makes small models useful as local assistants: following instructions, avoiding repetitive loops, calling tools on mobile devices, and giving useful answers without unnecessary output. The 9B model uses roughly 40% fewer tokens on standard instruction-following and everyday math tasks. Solving the Riemann hypothesis can wait.
These models are successors to our first-generation Bonsai 1.7B, 4B, and 8B, and provide a foundation for bringing practical agentic functionality to local applications on personal computers, phones, and wearables.
The Bonsai 2 family now includes four sizes spanning 2B to 27B parameters. Developers can choose the balance of capability, memory use, and responsiveness that fits their application and hardware.
Available today under Apache 2.0.
Harsh 💢 retweeted
Image-to-Pixel 3.0 is out 🎨
Turn images and video into pixel art right in your browser. Free, MIT open source.
Try it: tezumie.github.io/Image-to-P…
Or use it in code: npm i image-to-pixel
Harsh 💢 retweeted
"I don't have a GPU" is officially dead 🤯
You can now run 70B model on a single 4GB GPU and it even scales up to the colossal Llama 3.1 405B on just 8GB of VRAM.
AirLLM uses "Layer-wise Inference." Instead of loading the whole model, it loads, computes, and flushes one layer at a time
→ No quantization needed by default
→ Supports Llama, Qwen, and Mistral
→ Works on Linux, Windows, and macOS
100% Open Source.
Readers added context they thought people might want to know
AirLLM enables the claimed low VRAM usage via per-layer disk streaming but requires reloading the full model from storage for each token, yielding speeds like ~20s/token for 8B models and minutes per token for 70B+ per tests and project reports.
kunalvohra.com/blog/airllm-on…
github.com/lyogavin/airll…
Harsh 💢 retweeted
OCR is dead 🪦 Long live agentic OCR!
Traditional OCR makes one pass and hands back whatever text it got. Tables get flattened, charts disappear, and multi-column layouts come out scrambled. Nothing checks the output.
Agentic OCR treats parsing as a loop instead:
✅ layout-aware reading order
✅ smart routing of hard elements to the right model
✅ multi-pass verification and self-correction
✅ multimodal parsing of charts, images, and complex tables
Logan Markewich, LlamaIndex's Head of Open Source, wrote about what that shift looks like in practice and where it's still struggling. Link to the breakdown in the comments below.
Harsh 💢 retweeted
ASCII Magic now has an MCP server.
Ask Claude, Cursor or any MCP client for a look, and it renders through the same engine as the editor.
ascii-magic.com/mcp-server
Harsh 💢 retweeted
Introducing Kardashev-0.7, the world’s first trained swarm made of 32 distinct models
Trained with RL for Population Scaling (RLPS), 32 models organically develop specialization & complementary capabilities, delivering frontier performance at:
- 0.007x ~ 0.02x of the inference cost
- 0.03x of the required memory
Civilization advances through different minds specializing and working together. We’re bringing that principle into AI
@BanburyRoadAI, we’re scaling intelligence by model count, toward civilizations of models that learn to build on one another
Harsh 💢 retweeted
LLMs don’t need retraining to become more capable.
Our new architecture, Spotlight, gives AI models growing memory, allowing them to gain knowledge and capabilities without changing their weights.
Token generation takes constant work, no matter how much information is stored.
Harsh 💢 retweeted
Three weeks ago, we released the first version of Edge0:
“A 35B language model running on a local device, using only 1–2.5 GB of peak memory.”
Today, we’re taking the next step.
We’re open-sourcing the Edge0 app layer—now available on:
📱 iOS
🖥️ macOS
🤖 Android
🪟 Windows
One app. Four platforms. Fully on-device.
Harsh 💢 retweeted
We are launching Open Dots, run the same capabilities in 1/10th of the cost
OpenAI shipped Dot yesterday for Pro users.
Ours is open source, runs on your Mac and cheaper
Here's how it works:
- each dot has its own browser that stays logged in securely
- it can use and setup triggers on your Gmail, Calendar and 1,500 other apps
- you can call it and give it things to do while you talk
- dots can coordinate and schedule reminders
built with @composio @OpenRouter
free and open source: github.com/composio-communit…
Harsh 💢 retweeted
Introducing Ideogram 4.5, the most precise edit model.
With each edit, leading models add artifacts, pixel shifts, and color changes. Ideogram 4.5 eliminates artifact buildup, making multi-turn editing possible.
Live in Ideogram, the API, and launch partners. Open weights soon.
Harsh 💢 retweeted
for those that don't grasp how crazy this is
GPT Image 2.5 Sunburst vs Ideogram 4.5
and by the way it's likely just simple masking/in-painting but it still surprises me it's not the default for the majority of image editing models.
you can tell it's in-painting because you can see the bush around my shoulder slowly turn into a line. it's the same reason my tattoos fade on why side of my arm (the closest to my t-shirt).
no excuse for chopping of half of my head though.
I'm going to have to try a workflow where I use this on every frame of a video.
Harsh 💢 retweeted
Introducing jevgrep - a research agent CLI powered by jev from @typesafeai that reduces your coding agent cost by 40% (verified on SWE-bench)
Make sure to use the built in skill so your coding agent knows to use jg for context collection github.com/dzhng/jevgrep