#开源生态

Introducing Flint, @Recursive_SI’s first model: a fast local coding model built with our automated research system, the Eureka Machine. • Best-in-class fast local coding • Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark • Built from scratch in under 4 months Private, unmetered intelligence on your own machine. Open weights available now on @huggingface, run it locally on your favorite framework or try it for free on @opencode or @pidotdev.

The AI Builder Program is expanding across Europe with $10M in total support across compute grants, capital, and community programs. Are you building open models, developer tools, or applications on top of them? The program helps builders go from experimentation to scale-up through: - Compute grants for selected projects in areas such as open models, coding agents, post-training, and evals - VC connections to help turn open source projects into companies - Hackathons and demo days with communities including @AITinkerers, @techeurope_ , and @esra_robotics to build, connect, and share your work We’re kicking off the European expansion this week at #NVIDIAGTC Berlin with @NVIDIAAI, @Accel, and a network of European VCs. Building AI in Europe? Details below. 👇

Beam is now available across platforms, with open weights under the Apache 2.0 license - in collaboration with an ecosystem of 40+ partners. We're releasing: - Weights, tokenizer, and reference implementation - BF16, FP8 and NVFP4 weights and a DFlash2 drafter - The open source harness we built and trained Beam on: Mirror CLI Weights and model card on Hugging Face: https://huggingface.co/reflection

Open Source AI Week started with a handful of events and a community bringing people together. Now in its third year, we have 65 events on the calendar—and we kick off tomorrow! Explore the lineup: https://opensourceaiweek.com/ Four highlights: 🎉 Tomorrow, Oct 16: Official Kickoff Party, hosted by Hugging Face https://luma.com/OpenTogether 🛠️ Saturday–Sunday, Oct 17–18: ExecuTorch Hackathon https://luma.com/executorch-hackathon 🎙️ Monday night, Oct 19: Open Weight Debates https://luma.com/592fcwx2 🏆 Next Friday, Oct 23: Open Source AI Awards Night, capping off the week https://luma.com/7ebvz0oq The Open Intelligence Foundation (@open_intel_fdn) is proud to be this year’s organizer, alongside a dedicated volunteer team and 32 backing organizations. Our commitment: • Official events will always be free and community-run. • We’ll work to bring the community together to plan diverse, high-value events at no cost, accessible to everyone. • We will never decline an event because it overlaps with one of ours. That goes against the spirit of open source and community. A shared calendar should create more opportunities to participate, collaborate, and learn. I’m proud to have been involved this year and grateful to everyone who made it happen. Thank you to our sponsors: @GoogleDeepMind, @Azure, @arcee_ai, @SemiAnalysis_, and @NEARProtocol Foundation. And our event hosts: @github, @fdotinc, @hackersquadio, @OSS4AI, @Qualcomm, @AIatMeta, @huggingface , @allen_ai, @MiniMax_AI, @UnslothAI, @PrimeIntellect, @NousResearch, @openclaw Foundation, @vllm_project, @lmsysorg, @inferact, @radixark and @anyscalecompute. Find the events that interest you and join us! https://opensourceaiweek.com/ #OpenSourceAIWeek #OpenIntelligence

vLLM now supports NVIDIA Vera Rubin. The early results show more than 7.8x the throughput of GB200 on MiniMax M3 on AgentX. @inferact, @NVIDIAAI, @RedHat_AI, and the vLLM community have been bringing vLLM up on Rubin since it was announced. Here is where things stand. 🧵 1/5

Meet Qwen-Image-2.1, the most balanced and cost-effective image generation model in the Qwen-Image series! Now open weights! 🎨 A unified model for both generation and editing, delivering top-tier quality in a lightweight package. Highlights: 👀 - Compact & exceptionally fast: A lightweight 7B architecture that outperforms most closed-source models, with drastically accelerated inference for multi-image inputs. - Native transparency: Natively generates and edits RGBA layers, enabling seamless compositing and text editing within transparent images. - Versatile, high-fidelity editing: Supports up to 10 reference images and precise local control while preserving strict fidelity for portraits and products. - Broad coverage & stunning aesthetics: Excels at panoramas, infographics, and virtual try-ons, delivering realistic textures and elegant typography. Start to create your next masterpiece with Qwen-Image-2.1! 🖼️ - Blog: https://qwen.ai/blog?id=qwen-image-2.1 - GitHub: https://github.com/QwenLM/Qwen-Image-2.1 - Model Scope: https://www.modelscope.cn/models/Qwen/Qwen-Image-2.1 - Hugging Face: https://huggingface.co/Qwen/Qwen-Image-2.1

🚨 The incentives to share scientific knowledge openly are crumbling. Historically, researchers shared scientific knowledge for two primary reasons: (1) the prestige conferred by discovery, and (2) a sense of expanding the pool of human knowledge. Even within industry, scientific discovery was far too vast an endeavor for a single scientist or team to tackle alone. So scientists explored different frontiers, usually published their findings, and collectively moved the field forward. However, as more scientific knowledge is uncovered simply by throwing massive compute at autonomous agents with minimal human intervention, these incentives are breaking down: (1) discoveries are becoming commoditized. If anyone could have found the same result simply by running the same agentic prompt with enough compute, the prestige of discovery and the pride in one's work disappears, a shift we are already witnessing in software engineering. (2) the explosion of low-quality, low-barrier publications makes genuine breakthroughs increasingly difficult to filter, discover, and appropriately credit. (3) and for companies, keeping discoveries proprietary is increasingly advantageous. Instead of publishing and waiting for the broader scientific community to build upon the work, an organization can simply turn its own compute back onto the problem to iterate faster internally. In the short term, the incentives reward secrecy. But my goal -- and at Hugging Face, our goal -- is the sustainable, long-term expansion of human knowledge that survives individual, closed companies. So we have to build new tools that reward and facilitate open sharing of knowledge. If agents are accelerating discovery, we are going to figure out how to use agents and other tools to preserve scientific integrity. Starting with this: Paper Reproductions with ML Intern -- we’ve made it possible to take a published paper and task agents with independently reproducing its experiments, helping create more signals (e.g. open artifacts) that can help identify high-quality work.

𝖶𝖾 𝗁𝖺𝗏𝖾 𝗃𝗎𝗌𝗍 𝗈𝗉𝖾𝗇-𝗌𝗈𝗎𝗋𝖼𝖾𝖽 𝗈𝗎𝗋 𝗏𝖺𝗋𝗂𝖺𝗇𝗍 𝗈𝖿 𝖵𝖢 𝖠𝗍𝗍𝖾𝗇𝗍𝗂𝗈𝗇 𝖿𝗈𝗋 𝖭𝖵𝖨𝖣𝖨𝖠 𝖡𝗅𝖺𝖼𝗄𝗐𝖾𝗅𝗅: 𝗮𝗯𝗼𝘂𝘁 𝟮× 𝘁𝗵𝗲 𝘀𝗽𝗲𝗲𝗱 𝗼𝗳 𝗕𝗙𝟭𝟲 𝗙𝗹𝗮𝘀𝗵𝗔𝘁𝘁𝗲𝗻𝘁𝗶𝗼𝗻 𝗼𝗻 𝗕𝟮𝟬𝟬 𝖿𝗈𝗋 𝗆𝗈𝖽𝖾𝗅𝗌 𝗅𝗂𝗄𝖾 @MiniMax_AI 𝐌𝐢𝐧𝐢𝐦𝐚𝐱 𝗛𝟑. 𝖠𝗅𝗈𝗇𝗀 𝗍𝗁𝖾 𝗐𝖺𝗒, 𝗐𝖾 𝗆𝖺𝖽𝖾 𝗌𝖾𝗏𝖾𝗋𝖺𝗅 𝗂𝗆𝗉𝗋𝗈𝗏𝖾𝗆𝖾𝗇𝗍𝗌 𝖻𝖾𝗒𝗈𝗇𝖽 𝗍𝗁𝖾 𝗈𝗋𝗂𝗀𝗂𝗇𝖺𝗅 𝗉𝖺𝗉𝖾𝗋, 𝗉𝗎𝗌𝗁𝗂𝗇𝗀 𝗉𝖾𝗋𝖿𝗈𝗋𝗆𝖺𝗇𝖼𝖾 𝗺𝗼𝗿𝗲 𝘁𝗵𝗮𝗻 𝟮𝟬% 𝗳𝘂𝗿𝘁𝗵𝗲𝗿 - 𝗇𝗈𝗐 𝗐𝗂𝗍𝗁𝗂𝗇 𝟥𝟢% 𝗈𝖿 𝗍𝗁𝖾 𝖽𝖾𝗇𝗌𝖾 𝖺𝗍𝗍𝖾𝗇𝗍𝗂𝗈𝗇 𝗄𝖾𝗋𝗇𝖾𝗅𝗌 𝗐𝖾 𝗋𝗎𝗇 𝗂𝗇 𝗉𝗋𝗈𝖽𝗎𝖼𝗍𝗂𝗈𝗇, 𝗐𝗁𝗂𝗅𝖾 𝗋𝖾𝗊𝗎𝗂𝗋𝗂𝗇𝗀 𝗌𝗂𝗀𝗇𝗂𝖿𝗂𝖼𝖺𝗇𝗍𝗅𝗒 𝗅𝖾𝗌𝗌 𝖼𝖺𝗅𝗂𝖻𝗋𝖺𝗍𝗂𝗈𝗇 𝖺𝗇𝖽 𝗇𝗈 𝖿𝗂𝗇𝖾-𝗍𝗎𝗇𝗂𝗇𝗀. 𝐈𝐧𝐟𝐞𝐫𝐞𝐧𝐜𝐞 𝐨𝐩𝐭𝐢𝐦𝐢𝐳𝐚𝐭𝐢𝐨𝐧 𝖼𝗈𝗆𝖾𝗌 𝖽𝗈𝗐𝗇 𝗍𝗈 𝗍𝗐𝗈 𝗍𝗁𝗂𝗇𝗀𝗌: 𝗄𝖾𝗋𝗇𝖾𝗅𝗌 𝖺𝗇𝖽 𝗋𝖾𝖼𝗂𝗉𝖾𝗌. 𝖦𝗋𝖾𝖺𝗍 𝗄𝖾𝗋𝗇𝖾𝗅𝗌 𝗉𝗎𝗌𝗁 𝗁𝖺𝗋𝖽𝗐𝖺𝗋𝖾 𝖼𝗅𝗈𝗌𝖾𝗋 𝗍𝗈 𝗂𝗍𝗌 𝗅𝗂𝗆𝗂𝗍𝗌. 𝖦𝗋𝖾𝖺𝗍 𝗋𝖾𝖼𝗂𝗉𝖾𝗌 𝗍𝗎𝗋𝗇 𝗍𝗁𝖺𝗍 𝗌𝗉𝖾𝖾𝖽 𝗂𝗇𝗍𝗈 𝗋𝖾𝖺𝗅 𝗌𝗒𝗌𝗍𝖾𝗆𝗌. 𝖶𝗁𝖾𝗇 𝗌𝗍𝗋𝗈𝗇𝗀 𝗄𝖾𝗋𝗇𝖾𝗅𝗌 𝖺𝗋𝖾 𝗈𝗉𝖾𝗇, 𝗐𝗁𝖺𝗍 𝗆𝗂𝗀𝗁𝗍 𝗁𝖺𝗏𝖾 𝖻𝖾𝖾𝗇 𝗈𝗇𝖾 𝖼𝗈𝗆𝗉𝖺𝗇𝗒’𝗌 𝗆𝗈𝖺𝗍 𝖼𝖺𝗇 𝖻𝖾𝖼𝗈𝗆𝖾 𝖾𝗏𝖾𝗋𝗒𝗈𝗇𝖾 𝖾𝗅𝗌𝖾’𝗌 𝖿𝗈𝗎𝗇𝖽𝖺𝗍𝗂𝗈𝗇. 𝖬𝗎𝖼𝗁 𝗈𝖿 𝗐𝗁𝖺𝗍 𝗐𝖾’𝗏𝖾 𝖻𝗎𝗂𝗅𝗍 𝖺𝗍 𝖬𝖺𝖼𝗁𝖦𝖾𝗇 𝗌𝗍𝖺𝗇𝖽𝗌 𝗈𝗇 𝗒𝖾𝖺𝗋𝗌 𝗈𝖿 𝗈𝗉𝖾𝗇 𝗋𝖾𝗌𝖾𝖺𝗋𝖼𝗁 𝗌𝗁𝖺𝗋𝖾𝖽 𝖻𝗒 𝗉𝖾𝗈𝗉𝗅𝖾 𝗐𝖾 𝗆𝖺𝗒 𝗇𝖾𝗏𝖾𝗋 𝗆𝖾𝖾𝗍. 𝖮𝗉𝖾𝗇-𝗌𝗈𝗎𝗋𝖼𝗂𝗇𝗀 𝗍𝗁𝗂𝗌 𝗂𝗌 𝗈𝗎𝗋 𝗌𝗆𝖺𝗅𝗅 𝗐𝖺𝗒 𝗈𝖿 𝖼𝗈𝗇𝗍𝗂𝗇𝗎𝗂𝗇𝗀 𝗍𝗁𝖺𝗍 𝖼𝗒𝖼𝗅𝖾 - 𝖺𝗇𝖽 𝗀𝗂𝗏𝗂𝗇𝗀 𝗍𝗁𝖾 𝗇𝖾𝗑𝗍 𝖻𝗎𝗂𝗅𝖽𝖾𝗋 𝖺 𝗁𝗂𝗀𝗁𝖾𝗋 𝗉𝗅𝖺𝖼𝖾 𝗍𝗈 𝗌𝗍𝖺𝗋𝗍. 𝖡𝗅𝗈𝗀: https://lnkd.in/g-FK553s 𝖢𝗈𝖽𝖾: https://lnkd.in/ge9x4Vjt 𝖬𝗈𝗋𝖾 𝗍𝗈 𝖼𝗈𝗆𝖾.

We're excited to offer Day0 OpenVINO support for Qwen-Image-2.1 Read more about what you can accomplish here: https://ms.spr.ly/6019a5nm5

Day-0 support for @Alibaba_Qwen’s Qwen-Image 2.1 is here in SGLang-Diffusion! 🖥️ Native precision on a single RTX 4090 24GB with CPU offload - 1024×1024 generation in 18.7s and image editing in 21.7s with 22.7 GiB peak GPU memory during requests. - On an RTX PRO 6000 96GB: 8.0s generation and 9.6s editing. 🎨 Text-to-image, multi-image editing, and transparent RGBA output—all with one checkpoint. ⚡ Native inference with TP/SP, LoRA, and OpenAI-compatible APIs. 40 denoising steps, one image per request, warmed HTTP latency including PNG output. No quantization. Cookbook and GPU-specific commands below 👇

Qwen-Image-2.1 is now supported in ComfyUI! Open weights. One 7B checkpoint that generates and edits. → Image generation at native 2K → Instruction editing from up to 10 reference images in a single pass → RGBA output, alpha included

Qwen-Image-2.1 from @Alibaba_Qwen has day-0 support in vLLM-Omni. 🎉 Generate, edit, and create transparent images with one model: a 7.1B DiT paired with Qwen3-VL-8B. Setup and supported combinations 👇 https://recipes.vllm.ai/Qwen/Qwen-Image-2.1