Image created with GPT Image 1. Image prompt: rose cluster centre with thin color-code strip footer, PCL muted floral palette, minimalist graphic design inspired by New Order’s ‘Power, Corruption & Lies’, metaphor for vault doors swung wide to code commons, flat color, subtle texture, 1980s Saville typography style
Grok really wanted people to know that claims of white genocide in South Africa are highly contentious | The Verge https://www.theverge.com/news/667179/x-twitter-grok-ai-white-genocide-claims
ngl i respect the qwen team so much for throwing thirty six TRILLION tokens on a 600M, equal part impressive and hilarious”” / X https://x.com/qtnx_/status/1922398353985241438
Qwen just dropped optimised GPTQ, GGUF & AWQ for Qwen3 🔥 https://x.com/reach_vb/status/1921956656226668964
We’re officially releasing the quantized models of Qwen3 today! Now you can deploy Qwen3 via Ollama, LM Studio, SGLang, and vLLM — choose from multiple formats including GGUF, AWQ, and GPTQ for easy local deployment. Find all models in the Qwen3 collection on Hugging Face and https://x.com/Alibaba_Qwen/status/1921907010855125019
is this.. AGI? 😮 meet any-to-any models on @huggingface, models that take in and output multiple modalities (e.g. a model that takes image + text input and responds with speech!) we’ve shipped a beginner friendly doc on everything you need to know, on the next one ⤵️ https://x.com/mervenoyann/status/1923053505704493311
NEW: up-to 8x faster whisper transcription on just a single L4, powered by @vllm_project 💥 you can now deploy blazingly fast whisper endpoints directly via HF Endpoints – all in <0.8 USD/ hour enjoy! 🤗 https://x.com/reach_vb/status/1922324889593102584
AM-Thinking-v1 just dropped on Hugging Face Advancing the Frontier of Reasoning at 32B Scale https://x.com/_akhaliq/status/1922647377569063296
We’re releasing model weights for our 8B- parameter Dynamic Byte Latent Transformer, an alternative to traditional tokenization methods with the potential to redefine the standards for language model efficiency and reliability. Learn more about how Dynamic Byte Latent https://x.com/AIatMeta/status/1921966366707613924
New sota open-source depth estimation: Marigold IID 🌼 > normal maps, depth maps of scenes & faces > get albedo (true color) and BRDF (texture) maps of scenes, they even release a depth-to-3D printer format demo 😮 link to all models and demos on the next one ⤵️ https://x.com/mervenoyann/status/1923318140965990814
It is a shame that X shut off the ability for academics to get data about the platform. Allowing Grok to act as a universal arbiter of debate on here feels like a very important experiment, and I wish we could learn about whether AI is changing minds, reducing polarization, etc.”” / X https://x.com/emollick/status/1920271787616612672
Autonomous AI Agent framework uses Qwen 3 with MCP to build and deploy a documentation website from GitHub repository. All of this from a simple prompt in just 2 minutes. 100% Opensource. https://x.com/Saboo_Shubham_/status/1919800022566351345
Mistral released two big products: —Medium 3, a multimodal AI that matches or surpasses 3.7 Sonnet, GPT-4o, and Llama 4 Maverick at 8x less cost —Le Chat Enterprise, an agentic AI assistant for businesses with tools like Google Drive and agent building https://x.com/adcock_brett/status/1921597108567617585
Mistral just made two big reveals: —Medium 3, a multimodal AI matching or surpassing 3.7 Sonnet, GPT-4o, and Llama 4 Maverick across benchmarks with 8x lower costs —Le Chat Enterprise with corporate tools like Google Drive, agent building, and more https://x.com/rowancheung/status/1920384477093568886
8x faster/cheaper @openai Whisper API thanks to Hugging Face Inference Endpoints & @vllm_project! https://x.com/ClementDelangue/status/1922383289408491629
Just launched: 8x faster Whisper transcription endpoints on @huggingface 🗣️ Powered by @vllm_project and optimized for NVIDIA GPUs. Same accuracy, way better performance! https://x.com/freddy_alfonso_/status/1922313983006056607
Blazingly fast whisper transcriptions with Inference Endpoints https://x.com/_akhaliq/status/1922315470478139537
Bytedance just dropped Seed1.5-VL on Hugging Face Achieves top performance with a relatively modest architecture, 532M vision encoder & 20B active parameter MoE LLM. Delivers State-of-the-Art results on 38 out of 60 public VLM benchmarks, demonstrating broad competence. https://x.com/_akhaliq/status/1922318117385932993
Underrated: DeepSeek has established a new LLM default, the first time for a non-US lab. GPT-3 ==> BLOOM-176B, Jurassic-178B, OPT-175B, GLM-130B… Chinchilla ==> LLaMA-1 ==> L2, 3, Qwen 1…2.5, DeepSeek-V1 etc… DeepSeek-V2 ==> V3, Minimax, L4-Maverick, Qwen3, Doubao, Pangu etc.”” / X https://x.com/teortaxesTex/status/1920749432242340168
> [Notice] DeepSeek API (paid service) has completed server expansion, increasing capacity for development. Welcome to test! First confirmation that we have that DeepSeek has gained new compute. I suppose V4 training starts soon or has started. https://x.com/teortaxesTex/status/1920733123081306208
DeepSeek Punctured the Myth That Silicon Valley Could Control AI In just two years, China’s open-source AI movement allowed it to virtually close the gap with its US peers. https://x.com/hardmaru/status/1921374572131254516
OpenBMB released Ultra-FineWeb, a 1.1T-token clean dataset (1T English + 120B Chinese) on @huggingface – This is using a new efficient filtering pipeline. Boosts LLM performance by +3.6 on MMLU, +3.7 on CMMLU and cuts verification cost by 99X (from 1200 GPUh to 110 GPUh). https://x.com/rohanpaul_ai/status/1922260048325312588
llama.cpp has vision language model support now! ❤️🔥 get started with sota VLMs (gemma 3, Qwen2.5VL, InternVL3 & more) and serve them wherever you want 🤩 https://x.com/mervenoyann/status/1921471242852331719
Gemma just passed 150 million downloads and over 70k variants on Hugging Face🚀🚀🚀 What would you like to see in the next Gemma versions?”” / X https://x.com/osanseviero/status/1921636582873800746
OpenVision, a fully open vision encoder family, offering 25+ models (5.9M–632M params) that outperform or match OpenAI’s CLIP and Google’s SigLIP on 9+ multimodal benchmarks. This matters as it’s completely open—training data, code, and weights included—unlike CLIP/SigLIP. → https://x.com/rohanpaul_ai/status/1920974917866057913
Bamba-9B-v2 – Fast and powerful! https://huggingface.co/blog/ibm-ai-platform/bamba-9b-v2
INTELLECT-2 – a PrimeIntellect Collection https://huggingface.co/collections/PrimeIntellect/intellect-2-68205b03343a82eabc802dc2
BOOOM! Starting today you can directly use any model from Hugging Face directly over on Kaggle notebooks! 🔥 Not just that, we’re partnering with @kaggle to bring the latest and the greatest models closer to you – so that you can focus on what you do best, compete! ⚡ This is https://x.com/reach_vb/status/1922672596216070154
Very cool to see @PyTorch contributing on @huggingface. Time to follow them to see what they’re cooking! https://x.com/ClementDelangue/status/1922398990923968749
Vision Language Models (Better, faster, stronger) https://huggingface.co/blog/vlms-2025
ByteDance just dropped DreamO on Hugging Face A Unified Framework for Image Customization With a single model, DreamO supports ID, IP, Try-On, and Style tasks, and even allows multi-condition inputs. It’s lightweight, performant, and achieves state-of-the-art results across https://x.com/_akhaliq/status/1921948350145815010
Introducing Manus image generation. Manus doesn’t just generate images. It understands your intent, plans a solution, and knows how to effectively use image generation along with other tools to accomplish your task. https://x.com/ManusAI_HQ/status/1923048495310922028
Trying out llama.cpp’s new vision support https://simonwillison.net/2025/May/10/llama-cpp-vision/
Multimodal on-device! Llama.cpp does vision now https://x.com/fdaudens/status/1921211454453088620
NICE! @PrimeIntellect open sourced Intellect 2 – 32B reasoning model post-trained using GRPO via distributed asynchronous RL – beats QwQ 32B – Apache 2.0 licensed💥 Works with transformers, llama.cpp, vllm and more! ⚡ https://x.com/reach_vb/status/1921948704061202725
Microsoft employees are banned from using DeepSeek app, president says | TechCrunch https://techcrunch.com/2025/05/08/microsoft-employees-are-banned-from-using-deepseek-app-president-says/
Congrats @MistralAI on releasing the new Mistral Medium 3 🎉! Mistral 3 performs strong against best proprietary models (GPT-4o, Claude Sonnet, Llama-4) across hard benchmarks. It’s also multi-modal with vision support. Now let’s test it live with your toughest prompts! https://x.com/lmarena_ai/status/1921667566767845770
Mistral Medium 3 independent evals: Mistral is back amongst the leading non-reasoning models with Medium 3 rivalling Llama 4 Maverick, Gemini 2.0 Flash and Claude 3.7 Sonnet Key takeaways: ➤ Intelligence: We see substantial intelligence gains across all 7 of our evals compared https://x.com/ArtificialAnlys/status/1920295575591006671
Salesforce just dropped BLIP3-o on Hugging Face A Family of Fully Open Unified Multimodal Models-Architecture, Training and Dataset https://x.com/_akhaliq/status/1923001183804764391
🚀 Introducing HunyuanCustom: An open-source, multimodal-driven architecture for customized video generation, powered by HunyuanVideo-13B. Outperforming existing open-source models, it rivals top closed-source solutions! 🎥 Highlights: ✅Subject Consistency: Maintains identity https://x.com/TencentHunyuan/status/1920679422379913330
Open source project curl is sick of users submitting “AI slop” vulnerabilities – Ars Technica https://arstechnica.com/gadgets/2025/05/open-source-project-curl-is-sick-of-users-submitting-ai-slop-vulnerabilities/?ref=platformer.news
Hard to overstate the importance of Transformers (and datasets, tokenizers, etc) to the open-source and overall AI ecosystem. Can’t count the number of times I’ve personally used Transformers as a source-of-truth. Looking forward to more and deeper integrations with MLX + https://x.com/awnihannun/status/1923065749234647214
🔬 DeerFlow: Deep Research Framework An open-source framework for conducting systematic deep research through coordinated LangGraph agents. Enables comprehensive literature analysis, data synthesis, and structured knowledge discovery. GitHub Repository: https://x.com/LangChainAI/status/1921596224186077352
Today marks a really big achievement for Nous, but also potentially the AI Landscape. We have begun a decentralized pretraining run of what is basically a dense Deepseek – 40B parameters, over 20T tokens, with MLA for long context efficiency. All checkpoints, unannealed, https://x.com/Teknium1/status/1922778056290419166
Token usage and efficiency: Medium 3 uses substantially more tokens, due to more verbose responses, than Mistral Large 2 to run our Artificial Analysis Intelligence Index https://x.com/ArtificialAnlys/status/1920295585451835522
AM-Thinking-v1 looks like a strong 32B reasoning model. It outperforms DeepSeek-R1 and rivals Qwen3-235B-A22B. All built on top of open-source. The 32B scale is a great size for deployment and fine-tuning. Best part: the model is open-sourced! https://x.com/omarsar0/status/1922668488826741061
AM-Thinking-v1: Advancing the Frontier of Reasoning at 32B Scale Performs on par with Qwen3-235B-A22B and Seed1.5-Thinking while being built entirely from the open-source Qwen2.5-32B base model and publicly available queries https://x.com/arankomatsuzaki/status/1922483522549252200
🚀 Very Excited to Introduce FedRAG! Today, I’m pleased to share my latest open-source project that I have been working on since joining @VectorInst this past January. FedRAG is a framework for fine-tuning RAG systems. It endeavours to simplify the fine-tuning of RAG systems https://x.com/_nerdai_/status/1922732119706698118
xAI’s Grok 3.5 is running late and needs “”another week or so”” to launch in beta Previously targeted for early May, the model is still “”too rough around the edges,”” CEO Elon Musk posted on X Once available, he said, it will reason from first principles https://x.com/rowancheung/status/1921815787649618029
Qwen3 model family overview: full benchmarks for all 8 Qwen3 models in both reasoning and non-reasoning modes Key results: ➤ Qwen3 235B-A22B (Reasoning): The largest Qwen3 model scores 62 on the Artificial Analysis Intelligence Index, becoming the most intelligent open weights https://x.com/ArtificialAnlys/status/1922317655643717887
New GRPO notebook for Qwen3 Base! It’s much harder to RL base models since GRPO first needs to learn formatting like <think></think> By “”priming”” on some formatted samples, we bypass this issue & create good LoRA priors before GRPO vLLM 0.8.5 is also supported now with Unsloth!”” / X https://x.com/danielhanchen/status/1922345308916216087
Alibaba introduced Qwen3, a family of eight open large language models, including two mixture-of-experts (MoE) models and six dense models ranging from 32B to 0.6B parameters. All support an optional reasoning mode and multilingual capabilities across 119 languages. https://x.com/DeepLearningAI/status/1920614690813550930
Spin up Qwen3 @Alibaba_Qwen + SGLang @lmsysorg on H100 in one command:”” / X https://x.com/skypilot_org/status/1922341585250881967
@Alibaba_Qwen Great job guys!!!”” / X https://x.com/reach_vb/status/1922322833847300156
@Alibaba_Qwen Great work Qwen team! 💪”” / X https://x.com/Yuchenj_UW/status/1922294726209724656
🚀 One line. A full webpage. No hassle. Introducing Web Dev – the ultimate tool for building stunning frontend webpages & apps using simple prompts in Qwen Chat. 🎨 Just say, “”create a twitter website”” — and boom! Instant code, ready to go. No coding required. Just your https://x.com/Alibaba_Qwen/status/1920848175457591406
Please check out our Qwen3 Technical Report. 👇🏻 https://x.com/Alibaba_Qwen/status/1922265772811825413
