☯︎

just human

[Guillaume Meyer @guillaumemeyer]

🧹watermarks-remover now supports watermarks from OpenAI and Gemini in addition to Claude.
https://t.co/OxSdnAjEGe

[Jesper Nissen @JespernissenSEO]

Remove watermarks from Claude..

This guy allready has a tool to remove watermarks from Claude, as well as Gemini and Openai.

This is why I love being on X. Smart people sharing coll stuff..

[Liquid AI @liquidai]

Today, we release LFM2.5-VL-3B, a lightweight vision-language model that reads screens, documents, and the physical world. It handles digital screens across mobile, web, and desktop, grounds objects to coordinates, reads text and charts, and calls tools from either text or image input.

Built on LFM2.5-2.6B base, with a SigLIP2 400M NaFlex vision encoder

> Pre-trained on ~34T tokens
> Vocab size: 128K
Comparable or better scores compared to models up to 2.6x its size:
> ScreenSpot-v2 80.7, ahead of Gemma-4-E4B at 51.2
> RealWorldQA 73.1, ahead of InternVL-3.5-4B at 67.7
> TextVQA 84.3, ahead of Qwen3.5-4B at 81.2
> RefCOCO-avg 87.9, up from 57.1 on LFM2-VL-3B
> ToolSandbox 59.5, up from 26.4 on LFM2-VL-3B

🧵

[Korobochka (コロボ) 🇦🇺✝️🇷🇺 @cirnosad]

Revelation of the Pyramids by Jacques Grimault and Patrice Pooyard. This has been almost wiped out of the internet and is impossible to find. YouTube has deleted it. Why?

Note that the Brian Cox English narration deliberately censored this entire documentary, which is still useful because you can tell exactly what the most important parts are. The Brian Cox narrated version has not been censored in sharp contrast and is even available on Apple TV.

Make sure you watch, save this, and share it. It will literally change your life and how you view the world forever.

[dodo-reach 🦤 @DODOREACH]

4 months of using @pidotdev

this is post 3 about my pi setup (extensions and custom stuff)

half stolen, remixed and with a few pieces i already shared

a thread (with gems) ↓

[Luckey Faraday @luckeyfaraday]

I put all my vibe coded games in one place.

62 games. 13 models. Everything playable in your browser.

Multiplayer games, demos, experiments, model tests.

You can now try anything you’ve seen me post, here 👇

[NVIDIA AI @NVIDIAAI]

Introducing NVIDIA Nemotron 3.5 Lightning⚡

An open 30B MoE model with 3B active parameters, built for always-on agents to complete high-volume, specialized tasks faster.

It delivers up to 4x the output speed of similar-sized models.

[Mia @MiaAI_lab]

BIG update for DeepSeek v4 Flash 0731 for 2x @NVIDIAAI DGX Sparks ✨

- optional (experimental) vision support
- long context decode tok/s increased substantially
- improved multi-turn / agent tool calling
- optional abliterated weights path
- lots of bug fixes

Vision support is experimental, still working to make it work flawlessly, but you're welcome to give it a try! See the Vision section in the README for more details.

Some of these changes were made possible thanks to PRs and issues submitted to the GitHub repo - thank you!

Get it here:
https://t.co/6HXa9pxqhj

[David Stout @Davidstout]

Today, we’re excited to open-source TwiL-LM3, the first formal reasoning model from the webAI Intelligence Lab.

At just 3 billion parameters, TwiL-LM3 outperforms OpenAI’s GPT-OSS-120B on 4 of 5 formal reasoning benchmarks while running efficiently on consumer hardware. That’s 40× fewer parameters, 2.6× faster inference, and state-of-the-art performance in the reasoning tasks that power reliable tool calling, code generation, structured outputs, and AI agents.

TwiL-LM3 was trained using webAI’s proprietary reasoning pipeline on webAI-owned, verified datasets—not scraped internet data. We believe better reasoning comes from better training pipelines and higher-quality data, not simply larger models. Our approach demonstrates that efficient models can rival—and in many cases surpass—models dozens of times their size.

Designed for the edge, TwiL-LM3 runs on hardware people already own—from a Raspberry Pi to an iPhone—bringing advanced reasoning to millions of devices without relying on the cloud.

This is our first open-source release from the webAI Intelligence Lab, and it’s only the beginning.

Proudly built in Austin, Texas.

Article: https://t.co/ESW3D89xNX

[Royce White 🇺🇸 @Highway_30]

Don't tell me how to be a Black woman! Taking hormones are optional. The idea that a Trans woman has to subject herself to BIG PHARMA, is just another way for cis normative people to turn us into a commodity. I’m an ALL NATURAL Black trans lesbian with a penile aesthetic.

Royce White for U.S. Senate MN 2026
WNBA #LetHerPlay

[Mia @MiaAI_lab]

Unsloth Desktop to run and train local models! 🔥

Definitely gonna try it

[David Hendrickson @TeksEdge]

🔥 A new "tiny" model is available for Local AI that beats OpenAI's GPT-OSS 120B (high) on AA.

It's small, but worthy of your attention for more than that, and here's why.

Ling-3.0-tiny is only 7.9B total / 1.3B active parameters, yet it's designed for reasoning, tools and real agentic work.

🧠 128-expert sparse MoE
⚡ Only 8 routed + 1 shared expert active/token
🧩 3:1 KDA + MLA hybrid attention
🤖 Thinking + tool calling + agents
🔓 MIT licensed / open weights

And it's really small:
💾 BF16: ~15.8GB
💾 FP8: ~8.4GB
💾 INT4: ~5.8GB

FP8 reportedly uses only ~8.34 GiB peak memory at 8K context.

Performance:
🚀 DGX Spark: ~100–105 tok/s 👈
🍎 M4 Pro MacBook: ~86–90 tok/s
🧠 GPQA Diamond: 73.4
📐 IMO-AnswerBench: 71.0

Runtime support:
✅ SGLang
✅ vLLM via Ling branch
🧪 Ollama + MLX on Apple Silicon
⏳ llama.cpp / LM Studio still need to catch up
📱 Could this eventually run on a phone?

The 5.8GB INT4 weights make that surprisingly plausible on high-memory phones, but runtime support isn't there yet.

🎯 1.3B active parameters is starting to deliver capabilities we'd normally associate with much larger models.

This is the direction I want Local AI to go, smaller active models, less memory, faster inference, but still useful enough to run agents.

[Youssof Al Toukhi @Youssofal_]

Pi beating bloated OMP is hilarious.

[Tencent Hy @TencentHunyuan]

Introducing Hy3D WorldClaw——an agentic workflow that generates large scale 3D open worlds from text prompts.🚀🚀🚀
Not video, Not Gaussian Splatting, Every scene generated by WorldClaw is freely explorable and built entirely from editable, game-ready 3D assets with high-quality geometry and textures.
Project Page: https://t.co/KJXFWHLb1S

[Cactus Compute @cactuscompute]

We release Needle 2: A 14MB agentic LLM for phones, wearables, smart home, robots and microcontroller. The whole model is a single 14MB binary that runs a full session in 28MB of RAM. It is built on our Simple Attention Network findings, compressed to CQ2-bit with Cactus Quants, and baked into its own engine.

Needle 2 has 45m parameters trained from the ground up on 140B tool call, device use structured generation tokens. On mobile device use benchmarks, Needle 2 trades wins with frontier small LLMs like LFM2.5 230M, Apple FM Gemma-270m, at 5× to 70× smaller, and 2 bits against their f16.

Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, between 400–1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges 300–700 on sub-$200 phones such as the Samsung A-Series. Needle also runs on newer microcontrollers like ESP32.

A conventional transformer of Needle's width and depth spends 164 MFLOPs per token, and even one squeezed down to Needle's parameter count spends 87, Needle spends 70. Even on a high-end phone, an always-on assistant lives inside a power budget; every MFLOP is milliwatt-hours, and Needle spends 7x to 85x fewer of them per token than the smallest performant LLMs.

Read more: https://t.co/dLK2tIKXQu

[Stagehand @Stagehanddev]

Introducing Stagehand v4: the SDK for browser agents.

Playwright was built for testing, we built Stagehand for your agent: with improved context management, self-healing actions, and iframe support.

[How To Prompt @HowToPrompt__]

NVIDIA built a AI that nukes 30 years of animation technology.

They open-sourced a model that generates 350,000 motion skills in real-time, at 15,000 fps with just 2ms latency.

It’s called MotionBricks, an AI framework built to generate real-time character and humanoid motion at scale.

→ Intuitive smart primitives for quick scene authoring.
→ Integrated into NVIDIA's GR00T robotics stack.
→ Zero mocap. Zero rigging.
→ Runs in real-time

[Applied Compute @appliedcompute]

You can now post-train a model inside your existing production harness with our platform, AC2.

A production harness is a whole engineered system around the LLM, with its own context management, tools, sandboxing, and control flows. Porting that into a new training runtime can be expensive and could introduce train-test mismatch, where the policy is optimized against a simulated harness and then struggles in production.

All you need to do is swap out the harness’ LLM response endpoint to one provided by AC2, and expose a lightweight protocol for AC2 to initiate and grade rollouts; the trainer handles the rest.

[Elaya @elayadesigns]

A few prompts + Opus 5 + Higgsfield
That's all it took to build this cinematic scrollytelling landing page.

UI design skills: https://t.co/Gs1RT1BZxi

[Meng To @MengTo]

I made a three.js landing page with 3D scrolling for every section of the site.

Since AI got so good with 3D, you can replace videos with 3D instead. Apart from images, the whole site is 922 KB on disk, 290 KB gzipped. That's insane when a scrolling video site runs 20-100mb at 1080p depending on length.

As usual I had to prompt for improved textures, lighting and alpha masking. I used Claude Code (desktop) with Opus 5, Higgsfield for images, and https://t.co/tECuh9VUFz for the cloth effect on them.

Inspiration was Daniel Snow's awesome landing page: https://t.co/nwYLieptXl

Let me know if I should open-source. Don't wanna share too many without people asking.