
I love Cycling, Machine Learning and Backend.

Machine Learning & Backend 😎 Founder of Humanity backed by Investors from Google, SF and Hong Kong.

I love Cycling, Machine Learning and Backend.

𝗟𝗼𝗼𝗽𝗙𝗼𝗿𝗴𝗲: 𝗦𝗲𝗹𝗳-𝗛𝗼𝘀𝘁𝗲𝗱 𝗣𝗹𝗮𝘁𝗳𝗼𝗿𝗺 𝗧𝗵𝗮𝘁 𝗧𝘂𝗿𝗻𝘀 𝗔𝗜 𝗖𝗼𝗱𝗶𝗻𝗴 𝗶𝗻𝘁𝗼 𝗥𝗲𝗹𝗶𝗮𝗯𝗹𝗲 𝗧𝗲𝗮𝗺 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀 Tencent open-sourced LoopForge a self-hosted, self-evolving cloud platform for AI software delivery. Command like “fix this bug” run properly is luck. LoopForge turns your team’s delivery method into SOPs that coding agents can loop-run in the cloud: • Roles, not vibes Plan, code, review, and test are separate agents • Script gates + human sign-off Builds and tests decide by exit code • Engineering-grade RSI Each run is scored and improvements feed back into SOPs • Bring your own agents Works with CodeBuddy, Codex, Claude Code, Cursor, and more MIT licensed. Self-hosted. Your keys and code stay on your infrastructure. GitHub: github.com/Tencent/LoopForge #AI #OpenSource #CodingAgents #LoopForge

𝗛𝘆𝟰 𝗙𝗹𝗮𝘀𝗵 𝗪𝗲𝗶𝗴𝗵𝘁𝘀 𝗔𝗿𝗲 𝗡𝗼𝘄 𝗢𝗽𝗲𝗻 Tencent has released the open weights for Hy4 Flash. Get Hy4 Flash, Hy4 Flash-FP8, and Hy4 Flash Base: GitHub: github.com/Tencent-Hunyuan/H… GGUF: huggingface.co/AngelSlim/Hy4… Hy4 weights are coming next. #AI #OpenSource #Tencent #Hy4
🔓Hy4 flash weights are now open Get Hy4 flash, Hy4 flash-FP8, and Hy4 flash Base: -GitHub:github.com/Tencent-Hunyuan/H… -GGUF:huggingface.co/AngelSlim/Hy4… 👀Hy4 weights next

Some of you have been asking very politely. Some of you have just been asking. GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world. openai.com/index/introducing…

Today, we release d1.2 and d1.2-ultra, the state-of-the-art omnimodal decision models. d1.2 and d1.2-ultra both can process text, images, and audio with new output capabilities. > #1 on the @ArtificialAnlys Decision Index (75) > Read a value from the input (text, number or date) > Find every passage of a kind in a text (spans) > Locate things in a picture (box for one, boxes for every instance) > Estimate a number (number with a range) > Pick every option that applies (choice) 1/ 🧵

𝗬𝗢𝗟𝗢𝟮𝟲: 𝗨𝗻𝗶𝗳𝗶𝗲𝗱 𝗥𝗲𝗮𝗹-𝗧𝗶𝗺𝗲 𝗘𝗻𝗱-𝘁𝗼-𝗘𝗻𝗱 𝗩𝗶𝘀𝗶𝗼𝗻 𝗠𝗼𝗱𝗲𝗹𝘀 𝗙𝗮𝘀𝘁𝗲𝗿, 𝗔𝗰𝗰𝘂𝗿𝗮𝘁𝗲, 𝗡𝗼 𝗡𝗠𝗦 Ultralytics YOLO26 is the new unified real-time vision family that handles detection, segmentation, pose, classification, and oriented detection in one pipeline. What makes it special: • Native end-to-end inference no Non-Maximum Suppression (NMS) needed • Removed Distribution Focal Loss for a lighter, simpler head • New training tricks (MuSGD, Progressive Loss, STAL) for better accuracy and faster training • Strong gains on small objects and across all tasks Achieves 40.9–57.5 mAP on COCO while staying very fast on T4 TensorRT. Code: github.com/ultralytics/ultra… #AI #ComputerVision #YOLO #RealTime

𝗙𝗿𝗲𝗲𝗧𝗼𝗸𝗲𝗻: 𝗥𝘂𝗻 𝗛𝘂𝗴𝗲 𝗔𝗜 𝗠𝗼𝗱𝗲𝗹𝘀 𝗼𝗻 𝗬𝗼𝘂𝗿 𝗢𝘄𝗻 𝗖𝗼𝗺𝗽𝘂𝘁𝗲𝗿, 𝗘𝘃𝗲𝗻 𝗮 𝗟𝗮𝗽𝘁𝗼𝗽 A new system that lets you run very large open AI models locally on normal home hardware. It can run a 35B model on an 8GB laptop GPU, a 284B model on a gaming desktop, and even a 753B model on a single workstation GPU all at usable speeds for coding and agent tasks. It smartly uses both CPU and GPU, and adapts to whatever memory your machine has available. Code: github.com/FlashML-org/FreeT… Download: flashml.ai #AI #OpenSource #LocalAI

𝗠𝗼𝘃𝗶𝗲𝗥𝗼𝗹𝗹: 𝗔𝗜 𝗧𝗵𝗮𝘁 𝗠𝗮𝗸𝗲𝘀 𝗟𝗼𝗻𝗴, 𝗖𝗼𝗻𝘁𝗶𝗻𝘂𝗼𝘂𝘀 𝗩𝗶𝗱𝗲𝗼𝘀 𝗪𝗶𝘁𝗵𝗼𝘂𝘁 𝗟𝗼𝘀𝗶𝗻𝗴 𝗤𝘂𝗮𝗹𝗶𝘁𝘆 A new framework from Tencent Hunyuan and Peking University for generating long streaming videos (up to a minute or more) while keeping quality high. It produces video in chunks, one after another, so it can keep going without waiting for the full video. Special techniques help it remember past scenes efficiently and stay consistent across different events. The distilled version scores highest among compared methods on long-video benchmarks. Project: Tencent-Hunyuan.github.io/Mo… Code: github.com/Tencent-Hunyuan/M… Hugging Face: huggingface.co/yunyangge/Mov… #AI #VideoGeneration #MovieRoll

𝗦𝗲𝗻𝘀𝗲𝗡𝗼𝘃𝗮-𝗟𝗼𝗼𝗽-𝟵𝗕: 𝗦𝗺𝗮𝗹𝗹 𝟵𝗕 𝗠𝗼𝗱𝗲𝗹 𝗧𝗵𝗮𝘁 𝗠𝗮𝘁𝗰𝗵𝗲𝘀 𝗕𝗶𝗴𝗴𝗲𝗿 𝗢𝗻𝗲𝘀 𝗯𝘆 𝗥𝗲𝘂𝘀𝗶𝗻𝗴 𝗟𝗮𝘆𝗲𝗿𝘀 A new open-source model with only about 9 billion parameters that performs close to much larger 27B models on many tasks. How it works: • It reuses a block of layers twice (Loop + mHC), so 32 unique layers run as 56 layer executions. • Supports text and images, up to 512K context. • Official 4-bit version is only ~4.9GB and can run with as little as ~9GB VRAM at 64K context. Big improvements over the earlier Preview version on coding and agent benchmarks. GitHub: github.com/OpenSenseNova/Sen… Hugging Face: huggingface.co/sensenova/Sen… Collection: huggingface.co/collections/s… #AI #OpenSource #SenseNova
#SenseNova-Loop-9B is officially here — now #OpenSource! Just 9B parameters, competitive with Qwen3.8-27B across multiple benchmarks. 🔹 Big gains over Preview: NL2Repo 19.91 → 41.36; AutomationBench 14.33 → 36.67; GAIA 37.86 → 69.90 🔹 Loop + mHC: 32 unique layers, 56 layer executions 🔹 Day-0 4-bit AWQ weights (~4.9GB), with VRAM needs as low as ~9 GiB at 64K context 🔹 Up to 512K context, with text and image inputs See what a 9B model can build: animated posters, interactive websites and 3D scenes. 💻 GitHub: github.com/OpenSenseNova/Sen… 🤗 HF: huggingface.co/collections/s… 👾 Discord: discord.com/invite/BuTXPHmQu…


𝗦𝗮𝗴𝗲 𝗦𝟮.𝟭 𝗙𝗹𝗮𝘀𝗵: 𝗡𝗲𝘄 𝗔𝗜 𝗠𝗼𝗱𝗲𝗹 𝗦𝘁𝗿𝗼𝗻𝗴 𝗮𝘁 𝗦𝗰𝗶𝗲𝗻𝗰𝗲, 𝗦𝗽𝗿𝗲𝗮𝗱𝘀𝗵𝗲𝗲𝘁𝘀, 𝗮𝗻𝗱 𝗖𝗼𝗺𝗽𝘂𝘁𝗲𝗿 𝗧𝗮𝘀𝗸𝘀 A large 320-billion-parameter AI model built for both everyday work and scientific tasks. It can understand molecules, DNA, and scientific measurements, and also perform real actions like searching, coding, and using software. It ranks first or near the top on several benchmarks for office work, machine learning tasks, spreadsheets, and computer control. Homepage: internlm.github.io/Sage-S2.1… GitHub: github.com/InternLM/Sage-S2.… Hugging Face: huggingface.co/internlm/Sage… #AI #Science #Agents

Introducing Qwen3.8-27B-pi-adapted 🤗 We adapted Qwen's model specifically for the pi coding agent using reinforcement learning and a LoRA adapter. The results, measured as pass@1: 🚀 SWE-bench Pro (Hard 51): 29.4% → 62.7% 💻 Terminal-Bench 2.1: 39.5% → 55.8% ⚡ 55–62% fewer tool calls per task This is the power of open models: you can adapt them to YOUR tools and YOUR use cases. Let’s build! 💛 Open weights: huggingface.co/harness-hub/Q… Methodology: huggingface.co/blog/harness-…

Teaching robots new tasks just got a lot faster. 🤖 With Gemini 4 Hydrogen and the new robotics plugin for Google @Antigravity, you can connect a supported arm, practice in @MuJoCo simulation, and run tasks in the real world using plain conversation.

𝗝𝗘𝗣𝗔𝟰𝗥: 𝗥𝗼𝗯𝗼𝘁𝘀 𝗧𝗵𝗮𝘁 𝗦𝗲𝗲 𝘁𝗵𝗲 𝗥𝗲𝗮𝗹 𝟯𝗗 𝗪𝗼𝗿𝗹𝗱 𝗜𝗻𝘀𝘁𝗲𝗮𝗱 𝗼𝗳 𝗙𝗹𝗮𝘁 𝗣𝗶𝗰𝘁𝘂𝗿𝗲𝘀 Most robot systems only look at regular photos, which are flat. JEPA4R looks at the actual 3D shape of the room using a depth camera (like how some phones measure distance). It can imagine what will happen next and even show that imagination as a 3D view. It performs better than photo-based systems on robot tasks. After training only in a computer simulation, it works on a real robot arm right away. Paper: arxiv.org/pdf/2611.19771 codev: github.com/AIGeeksGroup/JEPA… #AI #Robotics

𝗜𝗻𝗰𝗹𝘂𝘀𝗶𝗼𝗻𝗠𝗲𝗱: 𝗔 𝗡𝗲𝘄 𝗕𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸 𝗧𝗵𝗮𝘁 𝗧𝗲𝘀𝘁𝘀 𝗔𝗜 𝗳𝗼𝗿 𝗥𝗲𝗮𝗹 𝗜𝗻𝗰𝗹𝘂𝘀𝗶𝘃𝗲 𝗛𝗲𝗮𝗹𝘁𝗵𝗰𝗮𝗿𝗲 𝗡𝗲𝗲𝗱𝘀 Most medical AI benchmarks focus on a few tasks. InclusionMed organizes evaluation by who the AI is helping (patients, doctors, or hospital admins) and what skill is needed (knowledge, reasoning, or taking action). They carefully selected 5 benchmarks and tested frontier models. Overall scores ranged from 17.5 to 49.1 out of 100. Rankings change a lot depending on the role or task. Models also differ widely in cost for similar scores. The goal is better, more inclusive evaluation of medical AI. Paper: arXiv:2611.21218 Site: inclusionmed.ai #AI #Healthcare #MedicalAI #Benchmarks

𝗔𝗹𝗹𝗼𝗰𝗙𝗹𝗼𝘄: 𝗔 𝗠𝗼𝗱𝗲𝗹 𝗧𝗵𝗮𝘁 𝗧𝘂𝗿𝗻𝘀 𝗥𝗮𝘄 𝗠𝗮𝗿𝗸𝗲𝘁 𝗗𝗮𝘁𝗮 𝗗𝗶𝗿𝗲𝗰𝘁𝗹𝘆 𝗜𝗻𝘁𝗼 𝗣𝗼𝗿𝘁𝗳𝗼𝗹𝗶𝗼 𝗣𝗼𝘀𝗶𝘁𝗶𝗼𝗻𝘀 Most systems first predict prices or signals, then build a portfolio. AllocFlow skips that and generates the full set of long and short positions in one step from raw market data. It uses a frozen market encoder plus a flow model. First it learns from an existing strategy, then reinforcement learning improves the results using real portfolio rewards. On hourly crypto futures, this lifts annualized return from ~28% to ~51% and the Calmar ratio from 0.96 to 4.04. It also works well on later unseen 2026 data. Paper: arXiv:2611.19937 Project: allocflow.github.io #AI #Quant #Trading