10 Notable Open-Weight Models for 2026 | Jihoo Kim님이 토픽에 대해 올림
10 Notable Open-Weight Models for 2026
이 제목은 아래 게시물에서 AI가 요약한 것입니다.
The open-weight race in 2026 has been nonstop. Here are 10 notable models. ✦ 𝗤𝘄𝗲𝗻𝟯.𝟱-𝟯𝟵𝟳𝗕-𝗔𝟭𝟳𝗕 (Alibaba) https://lnkd.in/gB23_37c Alibaba’s new flagship folds vision and language into one 𝗲𝗮𝗿𝗹𝘆-𝗳𝘂𝘀𝗶𝗼𝗻 𝗺𝗼𝗱𝗲𝗹 — a 397B MoE (17B active) that reads 𝘁𝗲𝘅𝘁, 𝗶𝗺𝗮𝗴𝗲𝘀, 𝗮𝗻𝗱 𝘃𝗶𝗱𝗲𝗼 𝗻𝗮𝘁𝗶𝘃𝗲𝗹𝘆, with a 262K context that stretches to 1M. ✦ 𝗛𝘆𝟯 (Tencent Hunyuan) https://lnkd.in/gAdujnjP A lean 295B MoE (21B active) tuned for reasoning and agents, Tencent’s Hy3 𝗽𝘂𝗻𝗰𝗵𝗲𝘀 𝘄𝗲𝗹𝗹 𝗮𝗯𝗼𝘃𝗲 𝗶𝘁𝘀 𝘄𝗲𝗶𝗴𝗵𝘁, rivaling flagship open models several times its size, all at a 256K context. ✦ 𝗗𝗲𝗲𝗽𝗦𝗲𝗲𝗸-𝗩𝟰-𝗣𝗿𝗼 (DeepSeek) https://lnkd.in/guzCccjp A 1.6T-parameter MoE (49B active) built for 1M-token context, its 𝗵𝘆𝗯𝗿𝗶𝗱 𝗖𝗦𝗔+𝗛𝗖𝗔 𝗮𝘁𝘁𝗲𝗻𝘁𝗶𝗼𝗻 makes long-context inference remarkably cheap. ✦ 𝗠𝗶𝗻𝗶𝗠𝗮𝘅-𝗠𝟯 (MiniMax) https://lnkd.in/gJHzadPe Trained on text, image, and video from the very first step, this 428B MoE (~23B active) pairs 𝗻𝗮𝘁𝗶𝘃𝗲 𝗺𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹𝗶𝘁𝘆 with sparse attention that keeps its 1M context fast and affordable. ✦ 𝗚𝗟𝗠-𝟱.𝟮 (Z.ai) https://lnkd.in/g5engRhM Z.ai ‘s flagship for 𝗹𝗼𝗻𝗴-𝗵𝗼𝗿𝗶𝘇𝗼𝗻 𝘄𝗼𝗿𝗸, a 753B MoE that sustains a 𝘀𝗼𝗹𝗶𝗱 𝟭𝗠 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 and leans hard into coding. ✦ 𝗜𝗻𝗸𝗹𝗶𝗻𝗴 (Thinking Machines Lab) https://lnkd.in/g2yxtxKx The debut open model from Thinking Machines is a 975B MoE (41B active) that’s unusually broad in what it perceives — text, image, video, and even 𝗮𝘂𝗱𝗶𝗼 𝗶𝗻𝗽𝘂𝘁, which few open models handle. ✦ 𝗠𝗼𝘁𝗶𝗳-𝟯-𝗕𝗲𝘁𝗮 (Motif Technologies) https://lnkd.in/gbkbtsqt A 𝗳𝗿𝗼𝗺-𝘀𝗰𝗿𝗮𝘁𝗰𝗵, 𝗳𝘂𝗹𝗹𝘆 𝗶𝗻-𝗵𝗼𝘂𝘀𝗲 314B MoE (~13B active) with custom GDLA attention and 256K context, multilingual by design. ✦ 𝗟𝗮𝗴𝘂𝗻𝗮-𝗦-𝟮.𝟭 (Poolside AI) https://lnkd.in/gbHhX26M Poolside’s 𝗰𝗼𝗺𝗽𝗮𝗰𝘁 118B MoE (8B active) is a 𝗰𝗼𝗱𝗶𝗻𝗴 𝘀𝗽𝗲𝗰𝗶𝗮𝗹𝗶𝘀𝘁 built for long-horizon agentic work, weaving reasoning in between tool calls across a full 1M-token context. ✦ 𝗦𝗼𝗹𝗮𝗿-𝗢𝗽𝗲𝗻𝟮-𝟮𝟱𝟬𝗕 (Upstage) https://lnkd.in/gv-FyCkZ Upstage’s 250B MoE activates just 15B per token, using a hybrid linear+softmax stack to deliver 𝗹𝗮𝗿𝗴𝗲-𝗺𝗼𝗱𝗲𝗹 𝗾𝘂𝗮𝗹𝗶𝘁𝘆 𝗮𝘁 𝘀𝗺𝗮𝗹𝗹-𝗺𝗼𝗱𝗲𝗹 𝗰𝗼𝘀𝘁 — an 𝗮𝗴𝗲𝗻𝘁𝗶𝗰 𝘀𝗽𝗲𝗰𝗶𝗮𝗹𝗶𝘀𝘁 with 1M context. ✦ 𝗞𝗶𝗺𝗶-𝗞𝟯 (Moonshot AI) https://lnkd.in/g6XUQfPD Moonshot’s Kimi-K3 is the 𝘄𝗼𝗿𝗹𝗱’𝘀 𝗳𝗶𝗿𝘀𝘁 𝗼𝗽𝗲𝗻 𝗺𝗼𝗱𝗲𝗹 𝗮𝘁 𝟯-𝘁𝗿𝗶𝗹𝗹𝗶𝗼𝗻-𝗽𝗮𝗿𝗮𝗺𝗲𝘁𝗲𝗿 𝘀𝗰𝗮𝗹𝗲, a natively multimodal (text + vision) MoE with 1M context and a highly sparse design. --- What they share: sparse MoE, long context (1M-token) — and most are free for commercial use. Which one are you most excited to try?
댓글을 보거나 남기려면 로그인
관련성이 더 높은 업데이트
- Great overview of how quickly the open-weight model market is advancing. We are proud to see Poolside’s Laguna S 2.1 included among the most notable open-weight models of 2026. What stands out about Laguna S 2.1 is the combination of intelligence, efficiency and control. With 118B total parameters and only 8B active, it is built for complex, long-running agentic software development while remaining small enough to deploy in environments customers control. For enterprise and government organizations, that means a high-performing, US-built model that can run on-premises, in secure or disconnected environments, without relying on a proprietary external model provider. The open-weight race is moving quickly, and Poolside is proud to be helping lead it. #OpenWeightAI #EnterpriseAI #AgenticAI #GovTech #SoftwareDevelopment #DOW #CDAO The open-weight race in 2026 has been nonstop. Here are 10 notable models. ✦ 𝗤𝘄𝗲𝗻𝟯.𝟱-𝟯𝟵𝟳𝗕-𝗔𝟭𝟳𝗕 (Alibaba) https://lnkd.in/gB23_37c Alibaba’s new flagship folds vision and language into one 𝗲𝗮𝗿𝗹𝘆-𝗳𝘂𝘀𝗶𝗼𝗻 𝗺𝗼𝗱𝗲𝗹 — a 397B MoE (17B active) that reads 𝘁𝗲𝘅𝘁, 𝗶𝗺𝗮𝗴𝗲𝘀, 𝗮𝗻𝗱 𝘃𝗶𝗱𝗲𝗼 𝗻𝗮𝘁𝗶𝘃𝗲𝗹𝘆, with a 262K context that stretches to 1M. ✦ 𝗛𝘆𝟯 (Tencent Hunyuan) https://lnkd.in/gAdujnjP A lean 295B MoE (21B active) tuned for reasoning and agents, Tencent’s Hy3 𝗽𝘂𝗻𝗰𝗵𝗲𝘀 𝘄𝗲𝗹𝗹 𝗮𝗯𝗼𝘃𝗲 𝗶𝘁𝘀 𝘄𝗲𝗶𝗴𝗵𝘁, rivaling flagship open models several times its size, all at a 256K context. ✦ 𝗗𝗲𝗲𝗽𝗦𝗲𝗲𝗸-𝗩𝟰-𝗣𝗿𝗼 (DeepSeek) https://lnkd.in/guzCccjp A 1.6T-parameter MoE (49B active) built for 1M-token context, its 𝗵𝘆𝗯𝗿𝗶𝗱 𝗖𝗦𝗔+𝗛𝗖𝗔 𝗮𝘁𝘁𝗲𝗻𝘁𝗶𝗼𝗻 makes long-context inference remarkably cheap. ✦ 𝗠𝗶𝗻𝗶𝗠𝗮𝘅-𝗠𝟯 (MiniMax) https://lnkd.in/gJHzadPe Trained on text, image, and video from the very first step, this 428B MoE (~23B active) pairs 𝗻𝗮𝘁𝗶𝘃𝗲 𝗺𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹𝗶𝘁𝘆 with sparse attention that keeps its 1M context fast and affordable. ✦ 𝗚𝗟𝗠-𝟱.𝟮 (Z.ai) https://lnkd.in/g5engRhM Z.ai ‘s flagship for 𝗹𝗼𝗻𝗴-𝗵𝗼𝗿𝗶𝘇𝗼𝗻 𝘄𝗼𝗿𝗸, a 753B MoE that sustains a 𝘀𝗼𝗹𝗶𝗱 𝟭𝗠 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 and leans hard into coding. ✦ 𝗜𝗻𝗸𝗹𝗶𝗻𝗴 (Thinking Machines Lab) https://lnkd.in/g2yxtxKx The debut open model from Thinking Machines is a 975B MoE (41B active) that’s unusually broad in what it perceives — text, image, video, and even 𝗮𝘂𝗱𝗶𝗼 𝗶𝗻𝗽𝘂𝘁, which few open models handle. ✦ 𝗠𝗼𝘁𝗶𝗳-𝟯-𝗕𝗲𝘁𝗮 (Motif Technologies) https://lnkd.in/gbkbtsqt A 𝗳𝗿𝗼𝗺-𝘀𝗰𝗿𝗮𝘁𝗰𝗵, 𝗳𝘂𝗹𝗹𝘆 𝗶𝗻-𝗵𝗼𝘂𝘀𝗲 314B MoE (~13B active) with custom GDLA attention and 256K context, multilingual by design. ✦ 𝗟𝗮𝗴𝘂𝗻𝗮-𝗦-𝟮.𝟭 (Poolside AI) https://lnkd.in/gbHhX26M Poolside’s 𝗰𝗼𝗺𝗽𝗮𝗰𝘁 118B MoE (8B active) is a 𝗰𝗼𝗱𝗶𝗻𝗴 𝘀𝗽𝗲𝗰𝗶𝗮𝗹𝗶𝘀𝘁 built for long-horizon agentic work, weaving reasoning in between tool calls across a full 1M-token context. ✦ 𝗦𝗼𝗹𝗮𝗿-𝗢𝗽𝗲𝗻𝟮-𝟮𝟱𝟬𝗕 (Upstage) https://lnkd.in/gv-FyCkZ Upstage’s 250B MoE activates just 15B per token, using a hybrid linear+softmax stack to deliver 𝗹𝗮𝗿𝗴𝗲-𝗺𝗼𝗱𝗲𝗹 𝗾𝘂𝗮𝗹𝗶𝘁𝘆 𝗮𝘁 𝘀𝗺𝗮𝗹𝗹-𝗺𝗼𝗱𝗲𝗹 𝗰𝗼𝘀𝘁 — an 𝗮𝗴𝗲𝗻𝘁𝗶𝗰 𝘀𝗽𝗲𝗰𝗶𝗮𝗹𝗶𝘀𝘁 with 1M context. ✦ 𝗞𝗶𝗺𝗶-𝗞𝟯 (Moonshot AI) https://lnkd.in/g6XUQfPD Moonshot’s Kimi-K3 is the 𝘄𝗼𝗿𝗹𝗱’𝘀 𝗳𝗶𝗿𝘀𝘁 𝗼𝗽𝗲𝗻 𝗺𝗼𝗱𝗲𝗹 𝗮𝘁 𝟯-𝘁𝗿𝗶𝗹𝗹𝗶𝗼𝗻-𝗽𝗮𝗿𝗮𝗺𝗲𝘁𝗲𝗿 𝘀𝗰𝗮𝗹𝗲, a natively multimodal (text + vision) MoE with 1M context and a highly sparse design. --- What they share: sparse MoE, long context (1M-token) — and most are free for commercial use. Which one are you most excited to try? 댓글을 보거나 남기려면 로그인
- Open-Weight Models and the Math of Acceleration “Open-weight models are inherently decelerationist.” That’s from a viral thread this week by an OpenAI exec, reacting to Kimi K3 — Moonshot’s 2.8-trillion-parameter release, the largest open-weight model ever, now #3 on the Artificial Analysis Intelligence Index. His conclusion: open weights undermine the business of AI and end in “AI communism.” I think the math says otherwise. Revenue = Price × Quantity. Open weights crush P — DeepSeek now prices some tokens ~97% below closed equivalents. Deceleration only follows if demand is inelastic. It isn’t. Every collapse in the price of intelligence has produced an even larger explosion in volume. That’s Jevons’ paradox in action, and it’s why hyperscaler capex went from 600–725B guided for 2026 — up as much as 77% — during the strongest two years of open-model progress in history. The structural point: open models are substitutes for closed models, but complements to compute. Kimi K3 doesn’t shrink the GPU market. It runs on it. And “public good” isn’t ideology — it’s Samuelson’s definition. Model weights are non-rival and non-excludable by construction. GPS is a free, state-provided, global public good. It didn’t give us communism. It gave us Uber. My falsification test: show me two consecutive years of falling capex alongside rising open-model quality share, and I’ll update. So far, we have the exact inverse. What’s your read — does free intelligence shrink the market, or grow it? 댓글을 보거나 남기려면 로그인
- 𝗪𝗼𝗿𝗸𝗶𝗻𝗴 𝗼𝗻 𝘀𝗼𝗺𝗲𝘁𝗵𝗶𝗻𝗴 𝗲𝘅𝗰𝗶𝘁𝗶𝗻𝗴 𝗹𝗮𝘁𝗲𝗹𝘆. 🚀 Over the past few days, I have been building an AI pipeline that transforms a simple sketch into a realistic image and then converts that 2D image into a 3D model. One of the biggest challenges wasn’t writing the code. It was getting an 𝘂𝗹𝘁𝗿𝗮-𝗹𝗮𝗿𝗴𝗲 𝗔𝗜 𝗺𝗼𝗱𝗲𝗹 𝘄𝗶𝘁𝗵 𝗵𝘂𝗻𝗱𝗿𝗲𝗱𝘀 𝗼𝗳 𝗯𝗶𝗹𝗹𝗶𝗼𝗻𝘀 𝗼𝗳 𝗽𝗮𝗿𝗮𝗺𝗲𝘁𝗲𝗿𝘀 running reliably and integrating it into the workflow. After debugging, GPU optimization, and testing, I successfully got two versions of the 3D generation pipeline working and producing usable 3D assets. I used ngrok to expose the locally hosted API through a public HTTPS URL. I initially considered Cloudflare 𝗳𝗿𝗲𝗲 𝗤𝘂𝗶𝗰𝗸 𝗧𝘂𝗻𝗻𝗲𝗹, but its approximately 𝟭𝟬𝟬-𝘀𝗲𝗰𝗼𝗻𝗱 𝗿𝗲𝘀𝗽𝗼𝗻𝘀𝗲 𝘁𝗶𝗺𝗲𝗼𝘂𝘁 was too short for my workflow. A full 3D generation request, including background removal, diffusion, and volume decoding, can take around two to four minutes. Ngrok worked better for these longer-running requests. A huge shoutout to Kaggle for making GPU-based experimentation with large AI model weights much more accessible. This is one milestone in a much bigger project. The goal is to create a smooth workflow from sketch to image to an exportable 3D asset. What AI project are you currently working on? #ArtificialIntelligence #GenerativeAI #ComputerVision #3DGeneration #Kaggle 댓글을 보거나 남기려면 로그인
- A few days ago I set out to see how far AI could take a real software project — not a toy demo, a full open-world racing game with physics, a garage/upgrade system, and actual races. The result: Velocity Streets, built almost entirely using Claude (Fable 5, Sonnet 5, and Opus 4 — each for what it’s best at). What used to take months of dedicated game-dev work took days. This isn’t about replacing developers — it’s about what becomes possible when you know how to direct AI well: importing 3D models, structuring game logic, iterating on physics, all through prompting and review rather than writing every line by hand. Curious what others building with AI tools are seeing — is this matching your experience, or am I still early on the curve? Watch the whole video on Youtube https://lnkd.in/djKithjy #ArtificialIntelligence #ClaudeAI #SoftwareDevelopment #AIProductivity #FutureOfWork #GameDevelopment #Anthropic #Innovation #TechLeadership #BuildInPublic 댓글을 보거나 남기려면 로그인
- 🚀 Kimi K3 is now open-weight and it’s a big deal. Moonshot AI just released the weights for Kimi K3, the world’s first open-weight model to hit 2.8 trillion parameters; and it’s free to download. Here’s why this matters: 🧠 2.8T parameters with a Mixture-of-Experts architecture (896 experts, 16 active per token); the largest open model ever released ⚡ 6.3× faster decoding at million-token context lengths, thanks to a new attention mechanism called Kimi Delta Attention (KDA) 📄 1M token context window; text, image, and video input supported out of the box 🏆 #1 on Frontend Code Arena, BrowseComp (91.2 vs Claude’s 88.0 and GPT’s 90.4), and multiple other benchmarks at launch ☁️ Day-0 hosted access via Together AI and Modal; no multi-GPU provisioning needed to get started Why this release matters Most frontier-model discussions still focus on parameter counts. Kimi K3 feels different. The design philosophy is clearly centered around: 🔹 Long-horizon coding and repository understanding 🔹 Agentic systems executing hundreds or thousands of tool calls 🔹 Multimodal reasoning workflows 🔹 Knowledge-intensive research and complex task execution What’s particularly exciting is that Moonshot AI didn’t just release the model. They also open-sourced significant parts of the surrounding ecosystem, including attention kernels, MoE communication infrastructure, and tooling that enables these large-scale training and inference systems. The open-weight race is no longer just about who has the largest model. It’s becoming a competition of architectural innovation, inference efficiency, and agentic capability. And Kimi K3 may be one of the most technically ambitious open releases we’ve seen this year. The funny part — Everyone celebrates “open weights.” Then they realize: ⏩ The model is 2.8T parameters. ⏩ Their gaming PC has 24GB VRAM. ⏩ The model was trained on infrastructure that costs more than their house. ⏩ Their 4090 is about to become a very expensive Discord machine. My RTX 4090 saw the announcement and immediately started drafting its resignation letter. #AI #KimiK3 #MoonshotAI #OpenSourceAI #LLM #GenerativeAI #AgenticAI #MachineLearning #MoE #LongContext #ArtificialIntelligence #kimi 댓글을 보거나 남기려면 로그인
- Today I learned about fonts. I had this question in my mind for a few weeks now, but I saved it till the end of the work on the first game. I asked the AI: “Can you check if the fonts we use in the game, are they assumed to be on the computer/device or are they bundled into the game?” AI: “They’re not in the game. The fonts you’re using, though, are basically in every device. And when they’re not, they’re automatically switched to similar fonts.” Me: “What would it take to add the fonts to the game itself?” AI: “Not a lot. Fonts don’t weight much. But you don’t really want to do that because you’d have to license even the fonts everyone has on their devices. You’re not allowed to just add them to your game.” Me: “Gotcha.” Every day a lesson. 댓글 6 댓글을 보거나 남기려면 로그인
- 🚨 THE OPEN-WEIGHT AI RACE JUST GOT SERIOUS. Kimi K3. 2.8T-parameter-class model. ~1M-token context. Strong agentic + coding capabilities. And one comparison that made me stop: 💰 93 Reportedly, both models completed 7/12 tasks in a 12-case agentic evaluation. But here’s the real story: The future of AI won’t be: ❌ One model ❌ One API ❌ One-size-fits-all architecture It will be: 🔥 Multiple models 🔥 Intelligent routing 🔥 Agents 🔥 Tools 🔥 Memory 🔥 Evaluation The question is no longer just: “Who has the smartest AI?” It’s: “Who can turn intelligence into reliable work — at the lowest cost?” I went deep into Kimi K3’s benchmarks, economics, 1M context window, limitations, and what it means for developers and founders. 👇 Full breakdown: #KimiK3 #ArtificialIntelligence #AIAgents #LLM #OpenSourceAI 댓글을 보거나 남기려면 로그인
- Luis Cancilleri the International Tribunal for Sustainable AI Use has adjudicated and has confirmed this as a case of #AiTaskSlop. As the task concerned required significant computational resources (e.g. to process a large image dataset) and the final product is of minimal utility (e.g. telling the user what is in their wardrobe), a total of 20 penalty penalty points would be applied to your account. However, as you did not execute this task yourself and are merely reposing content generated by another user, no penalty points will be applied in this instance. Please be more mindful in your use of AI. GPT-5.6 Sol just launched, and someone already found the most personal use case for it I’ve seen this week. A builder named Thijs gave the model access to his own camera roll, had it extract every piece of clothing he owns from old photos, then used GPT’s image model to render new outfit combinations on himself. No custom app. No dataset. Just his own photos and a few prompts. What this signals for builders: You don’t need a “market” to justify a project. Your own life is enough data. The barrier was never the model. It’s deciding what’s worth pointing it at. The smallest personal projects often teach you more about a new model’s real capabilities than any demo does — before you ever put it in front of a client. I keep coming back to this with every new model release. The demos are impressive. But the projects that stick with me are the small, specific, slightly self-indulgent ones. What’s the smallest project you’ve been sitting on that only matters to you? #OpenAI #AI #Agents 댓글을 보거나 남기려면 로그인
- GPT-5.6 Sol just launched, and someone already found the most personal use case for it I’ve seen this week. A builder named Thijs gave the model access to his own camera roll, had it extract every piece of clothing he owns from old photos, then used GPT’s image model to render new outfit combinations on himself. No custom app. No dataset. Just his own photos and a few prompts. What this signals for builders: You don’t need a “market” to justify a project. Your own life is enough data. The barrier was never the model. It’s deciding what’s worth pointing it at. The smallest personal projects often teach you more about a new model’s real capabilities than any demo does — before you ever put it in front of a client. I keep coming back to this with every new model release. The demos are impressive. But the projects that stick with me are the small, specific, slightly self-indulgent ones. What’s the smallest project you’ve been sitting on that only matters to you? #OpenAI #AI #Agents 댓글 47 댓글을 보거나 남기려면 로그인
팔로워 18,021명
- Source: Unknown