(╥﹏╥) | samuu | α|Δ|ι

Hey guys & @askOkara, I'm joining your affiliate program! 🚀 I've been using Okara for a while for quick secret chat stuff with my AI buddy that I don't want others to find out about. I chose it because of the Secure Mode—it's end-to-end encrypted so I know my data is actually private. 😍 Join Okara AI now: okara.ai/?via=samunder

Can you explain why ReLU kills gradients, why GELU is in every transformer, why Softmax turns logits into probabilities? > Sigmoid > ReLU > Tanh > Softmax > LeakyReLU > GELU > Swish > ELU > SELU Every activation function you'll ever need, explained by implemention. Practice all of them on TensorTonic: tensortonic.com

Introducing Unsloth Studio ✨ A new open-source web UI to train and run LLMs. • Run models locally on Mac, Windows, Linux • Train 500+ models 2x faster with 70% less VRAM • Supports GGUF, vision, audio, embedding models • Auto-create datasets from PDF, CSV, DOCX • Self-healing tool calling and code execution • Compare models side by side + export to GGUF GitHub: github.com/unslothai/unsl… Blog and Guide: unsloth.ai/docs/new/studio Available now on Hugging Face, NVIDIA, Docker and Colab.

Banger paper from Zhipu bros

So apparently this is how you wear the AirPods Crazy, I didn't know about this

Meet Waifu Diffusion: a specialized AI that transforms text into anime-style artwork. This model has captured the community's imagination, generating stunning character art with just a few descriptive words. It's a creative powerhouse for anime fans and artists alike.

Today we launch Fish Audio S2, a new generation of expressive TTS with absurdly controllable emotion. - open-source - sub 150ms latency - multi-speaker in one pass Real freedom of speech starts now 👇

@SarvamForDevs add this to openrouter

Sarvam-105B is trending on Hugging Face 🚀

Today we're releasing our first open source TTS model, TADA! TADA (Text Audio Dual Alignment) is a speech-language model that generates text and audio in one synchronized stream to reduce token-level hallucinations and improve latency. This means: → Zero content hallucinations across 1,000+ test samples → 5x faster than similar-grade LLM-based TTS → Fits much longer audio: 2,048 tokens cover ~700 seconds with TADA vs. ~70 seconds in conventional systems → Free transcript alongside audio with no added latency

Small and medium sized models (such as Qwen3.5) are now good enough to reliably handle tools but also generate small code blocks and eve fine tune for function calling! All done locally or on your iPhone. 📲

What tools have you added to @lmstudio and your new Qwen3.5-9B? Try MCPs for BreaveSearch, Memory (Knowledge Grepah), Fetch (webpage scraper)

🚀 FlashMaskV4: Leveling up with FlashAttention-4 With @tri_dao officially releasing the FlashAttention-4 paper, we are thrilled to announce the continued evolution of FlashMaskV4! Building on our research FlashMask (arxiv.org/abs/2410.01359), we’ve integrated FA4’s core power to deliver the ultimate sweet spot between masking flexibility and hardware-limit throughput. 🔥 Why FlashMaskV4? 🔹FA4 Powered: Fully leverages the latest FlashAttention-4 kernels for next-gen efficiency. 🔹Column-wise Sparse Masking: Optimized support for diverse masks (Prefix LM Document, Share Question, etc.) across both FWD and BWD. 🔹Massive Speedups: Up to 2.9x faster in FWD and 1.6x in total compared to base FA4 mask_mod (8k seq). 🔹Long-Context Mastery: Maintains high efficiency and stability from 8k to 128k sequence lengths. No more choosing between custom attention logic and peak FLOPS. ⚡️ Explore the code & benchmarks: 🔗 github.com/PaddlePaddle/f… #FlashAttention4 #FlashMaskV4 #MachineLearning #OpenSourceAI #LLM #PaddlePaddle

@samirande_ M*th marna kam kardo

Am I the only mf who has short term memory at the age of fkn 18

@techdevnotes Humanity ❌ gooners ✅

Grok is the only AI that promotes Humanity ❤️

Introducing Uni-1, Luma’s first unified understanding and generation model, our next step on the path towards unified general intelligence. lumalabs.ai/uni-1

@noobKoderr It's for free users , 5.4 for paid 👀

Didn't they launch GPT 5.3 a few days ago?

🚨 Someone just open sourced their own Neuro-sama. And it might be better than the original. It's called AIRI. A fully autonomous AI companion that talks to you in real time, plays Minecraft with you, plays Factorio, chats on Discord and Telegram, has a Live2D and VRM avatar body, remembers past conversations, and runs entirely on your machine. Not a chatbot. Not a voice assistant. A digital being that lives on your screen. Here's what this thing can actually do: → Real-time voice conversations with speech recognition + synthesis → Plays Minecraft autonomously alongside you → Plays Factorio (builds, automates, strategizes) → Joins your Discord and Telegram to chat → Full animated Live2D and VRM avatar with auto-blink, eye tracking, and idle animations → Persistent memory across sessions → Local inference via WebGPU. No API calls needed → Runs in your browser, on desktop (macOS/Windows), or on mobile Here's why this matters: Neuro-sama is the most famous AI VTuber on the internet. Millions watch her streams. But she's closed source. When the stream ends, she's gone. You can't talk to her. You can't play games with her. You can't customize her. AIRI is the open source answer. Your own AI companion. Always available. Fully yours. Here's the wildest part: It supports 30+ LLM providers. OpenAI, Claude, Gemini, DeepSeek, Ollama, Groq, Mistral, xAI, local models. Swap the brain with a config change. It uses native CUDA and Apple Metal for inference. Not some slow wrapper. Real GPU acceleration through HuggingFace Candle. It even has a plugin system so the community can extend what she can do. Watch videos. Browse the web. Learn new games. The architecture is built for it. 17.5K GitHub stars. 1.7K forks. 2,518 commits. 101 contributors. 46 releases. This isn't a weekend project. This is a movement. 100% Open Source. MIT License.

The model performance comparison between Nanbeige4.1-3B and Qwen3.5-4B (from its model card).

🚀Trillion parameters. Zero compromises. 100% open source. 🔥Introducing Yuan 3.0 Ultra — our flagship multimodal MoE foundation model, built for stronger intelligence and unrivaled efficiency. ✅️Efficiency Redefined: 1010B total / 68.8B activated params. Our groundbreaking LAEP (Layer-Adaptive Expert Pruning) algorithm cuts model size by 33.3% and lifts pre-training efficiency by 49%. ✅️Smarter, Not Longer Thinking: RIRM mechanism curbs AI "overthinking" — fast, concise reasoning for simple tasks, full depth for complex challenges. ✅️Enterprise-Grade Agent Engine: SOTA performance on RAG & MRAG, complex document/table understanding, multi-step tool calling & Text2SQL, purpose-built for real-world business deployment. 📂Full weights (16bit/4bit), code, technical report & training details — all free for the community. 👉Learn More: github.com/Yuan-lab-LLM/Y…

@lukaxko @xeophon You tried,/no_think

@xeophon How did you disable thinking? For some reason gguf unsloth q4 version does not produce thinking traces while mlx version always thinks no matter what I prompt

qwen3.5 4b + disabled thinking is so good, man

I built a neural network from scratch without using PyTorch, TensorFlow, or any libraries for that matter. Instead, I implemented the core math myself. I'm working on making my own machine learning framework from scratch with my friend @_reesechong. He previously trained a similar neural network, but using just scalars. The next step was to use tensors instead. The benefit of this is clear: when using scalars each data point is looped through separately creating its own node in a computation graph. With tensors, these separate nodes are stored together, allowing one forward and backward pass for the whole batch, greatly improving the efficiency of training. We often hear the saying "don't reinvent the wheel", but in my experience rebuilding technologies that abstract away a lot of complexity gives you a better and more thorough understanding of how the system works. Results of the training are shown below. Feel free to checkout the repo and read through the code, linked in the replies.