Marco (🫰,✨)

Yeah on dgx i can't run vllm with mtp (it runs but performance goes down 75% for some reason). I use sglang because of its approach with cache and mtp i can get to 14 tps on average with same ram i had with ollama. I think the benefit for you, if you use agents, is that the reasoning time is much lower: the non opus version it happens that loops on its own and most of the tok/sec goes waste on useless thinking.

@croll83 @0xSero Thanks Marco. NVFP4 being around 10GB larger than the 16GB Q4 GGUF is still a bit scary even on a 5090. Not sure if I want to go through all that trouble yet

Qwen3.5-35B compressed 20% with 1%~ performance drop on average. Now you can fit this (4bits) with full context on 24GB of VRAM 700$~ or 1x 3090 huggingface.co/0xSero/Qwen-3.…

I fully agree, i'm running on dgx and under stress with 128k context it takes 50/60 GB (no parallel requests). But i had to stress the stack because of the low ram bandwidth so i had to waste more ram to improve speed (i can get up to 14tok/sec, no more). I think this specific model is the best to use in everyday life with agents (in parallel to an opus subscription for real coding and deep clever stuff). No competition with other models, specially MoE. Original qwen 27B is the best for tool calling, but it lose the track with the reasoning. The opus distill make it perfect. Having that on a dedicated GPU possibly to 30/50 tok/sec would be a game changer.

Yeah, There is a very big missing info in your thread (vibecoding a miniapp it's not enough to make 1m$) and that makes all the difference in the world to target the right people with your very important core message (it's easiee to build on ton with ai and appkit, agent kit). I just would love to see more talented people joining TON and building a solid ecosystem. Clickbaits, false promises, bad execution does not help with this goal, and we had a lot so far, it's time to shift the approach.

Ofc to some degree this is a click-bait But I disagree with you, I think that you can vibecode an amazing product too, and in fact there are examples of it, and they are mentioned in the thread. The whole point is to make developing & building fun and accessible to an average person. So I don’t find it misleading as you suggest, but I see your pov of worry for an inflow of the low-quality MiniApps. Ofc you have to figure out virility, distribution, KOLs, which comes as a part of building a business. But Agentic AI already today can do a big part of it, such as build you funnels, apps, website, run your social pages & etc. IF the question is: can you vibecode a functioning product that would be generating revenue? — absolutely!

IMPORTANT: How to make your first $1,000,000 with OpenClaw AI Agent 🦞💎 - How to create your own business & scale - Build your own Telegram MiniApp & get millions of users - Examples of businesses built on Telegram by TON Community bookmark this post & let’s get into it 🧵✍️

no brother i disagree with the claim and the clickbait approach. of course you can create a business that makes 1M$. elon musk created a business that made multiple B$. The fact that notcoin made a multi-M$ business on TON does not imply that anyone else will do it just because they vibecode a miniapp. making 1M$ is the consequence of multiple factors. vibecoding the TMA is 0.1% of them. Probably the most important factor is getting virality and proper execution capability, which openclaw can't do shit about. Your post have an important message: building on ton is now easier thanks to appkit (which i really hope soon will add TAC methods to make it even richer) and the availability of a dedicated skill (Agent Kit). but framing this very important message in the context of 'make your first 1M$' is either a clickbait or the post is made to capture the attention of very low iq people ("hey claude, build something to make me rich"). In both cases it's not a good target. Real target is talented builders and founders and if they read this post, it's very probable they will not consider it serious with such an opening claim. And if we keep going this way we'll find hard times to capture the attention of talents on TON and we'll end up with 50M vibecoded miniapp form very low iq people with no chance of success. The only outcome will be to pump fake metrics of 'number of miniapps' - which historically has been a big mistake in marketing ton blockchain.

Well, you need to create a business & monetise it, and in fact I’ve provided examples of what ppl have built with Agentic AI. There are numerous projects who made millions of $ on TON & Telegram, in gaming/DeFi/NFTs This thread points you toward the direction to get started building something of your own, So what’s wrong? You disagree that businesses can generate over $1M? If so — here’s the thread with the data from 2025

@sekizaltmis_ @0xSero use google colab with llm-compressor, will take <5$ to use an A-100 instance and it will take 30min. i'm actually doing the FP8 to run with sglang right now locally on my DGX

@0xSero Need AWQ fot his repo huggingface.co/Jackrong/Qwen3…

@TheAhmadOsman wait how much power does this monster drain? is it a blackhole cluster?

New Tenstorrent cluster hot from the kitchen > 1TB of VRAM > 3TB DDR5 RAM > 32TB SSD Storage New product, will share more later P.S. Can you find the cat in the picture?

Milady

@MacroMate8 @SolanaFndn Congrats sir. Make the cake bigger so anyone will get a better slice of it 😍

big personal news: first off, Ethereum is dead second, i am now Special Situations at @SolanaFndn. my mandate is mega deals that will drive max value why solana? > the most engaged userbase in crypto with the highest take rate per volume (more than hype coinbase binance) > foundation is very lean and full of killers > other chains dont need me. ethereum is fucked (stuck in communism) and hype’s perp game is mature. solana is perfect for me to have fun and get shit done what will i do at solana: > bring 10B TVL to solana i’ve brokered 1-2B before and wanna up my game. if you are a large TVL provider, we have super scalable fixed yield products at 7-8% in a tradfi setup dm me > help grow RWA perps on solana hype done a great job with oil, gold, nasdaq perps and we will take it further if you are a serious crypto/tradfi market maker in rwa, we have the distribution and open to deals dm me what motivates me personally: first off, I am doing this for glory. I want to come in and be instrumental in growing trading activity and onchain tvl. second, I want to be relevant again. I miss pissing people off while being unavoidable. its fun thirdly, learn how to acquire trading distribution. almost no one in crypto knows how to do but thats where billion usd opportunities lie and they belong on solana.

@tayvano_ @_Enoch Oh that reminds me so much about the hundreds of useless RAPIDs I was forced to write over 3 years 🤣🤣🤣

Whatever this shit once was it’s now an echo chamber of idealistic virtue signaling and showcasing in-group membership above all else. This is not how you build a better anything. It’s how you form an increasingly exclusive cult that has increasingly less influence and achieves less. When you prioritize agreement and superficial signals over fucking results you fail to achieve anything of value. It’s not about money or what you think it’s valuable, it’s about doing things. And you can’t do things when you’re too fucking busy making sure that you’re all in fucking agreement about doing things. Grow the actual fuck up this is deeply fucking embarrassing.

It's fine; nobody is forced to like Milady. Anyone who wants to can post-author a version of the Mandate without her aesthetics. The text still stands. CROPS isn't pandering.

@waleswoosh @OnlyZen their success does not depend on launching the token in a super bear market to give you back your ICO money...

@OnlyZen My bags depend on their success

If all 53,000 MegaETH ICO participants bridged just $10 USD to MegaETH, we'd hit the $500M USDM KPI today. And TGE would be next week. We can do this

@0xCircusLover Long life to Tron, Algorand, ICP, and all the others 2018/2019 L1s still with B$ FDV 😅

-Farcaster -DoubleZero -Drift -Abstract (clear scam) -Aztec -Lighter -Avantis -90% of Monad end Megaeth ecosystem. -Polygon -LayerZero Did we miss something?

These brypto projects will be dead in 1-2 years: -Tempo -OP -Monad -Arb -Megaeth -Berachain (already ded) -Linea -Espresso -Zircuit -Axelar -0G -Story -Camp -Blast (already ded) -Aptos -Sui -Sei -Aleo -Base -Celestia -Eigenlayer -Avail -Scroll -Zksync -Starknet -Eclipse -Arcium

@MichaelAntonF @TeksEdge @AMD @Apple Which are the one you use? Llama.cpp? Tried vllm it takes 1.4x ram and was slower. Don't need multitenancy that's why opted for ollama

The Ultimate 128GB Local AI Hardware Battle 🥊💻 Judging Qwen3.5-27B (Bartowski IQ4_NL) on top unified-memory machines: 1️⃣ @AMD Strix Halo (Ryzen AI Max+ 395)💰 ~$2,500 | 🚀 9 - 12 tps (decode) | 🎮 Full Windows AAA gaming 🏆 Speed + value king. 🖕 2️⃣ @Apple Mac Studio M3 Ultra💰 ~$5,000 | 🚀 8–12 tps | 🍎 Apple macOS & ecosystem w/solid speeds; limited AAA gaming 3️⃣ @NVIDIA DGX Spark (GB10 Blackwell)💰 $4,699 | 🚀 ~10 tps (~20 tps x2 node) | 🐧 Linux/AI research only w/strong prefill + nerfed decode bandwidth-limited. Difficult pooling (new cables may fix). AAA gaming not optimized for Grace Verdict: AMD wins for most power users and best speed/price/gaming combo. (Community benchmarks; YMMV with setup/context) Which are you buying? 👇

My net worth peaked at $1.2 million. None of it was real. I don't mean that philosophically. I mean it was located on servers that have since been turned off. I own eleven properties in the metaverse. Three in Decentraland. Four in The Sandbox. Two in Voxels. One in Otherside. And a beachfront villa in Horizon Worlds that I bought for $214,000 because Mark Zuckerberg called it "the next frontier." The frontier closed last week. It's a mobile app now. Last year I mass DM'd 340 people the phrase "you don't understand how early we are." I have since stopped doing that. Not because I was wrong. Because most of them blocked me. I got into metaverse real estate in November 2021. Everyone was buying. Someone paid $450,000 to be Snoop Dogg's neighbor. In a video game. With no legs. The avatars didn't have legs. I thought that was bullish. "The legs are coming," I told my Discord. "Legs are a roadmap item." Three hundred people reacted with rocket emojis. I called myself a "digital land baron." I put it in my Twitter bio. I put it in my LinkedIn headline. I said it on a podcast that had eleven listeners. Three of them were bots. The rest were my alts. My virtual property has more square footage than my actual apartment. My actual apartment has furniture. Location, location, location. My most valuable asset was a plot next to a virtual Gucci store. Gucci left in 2023. The store is still there. Nobody's in it. It's like a mall in Ohio but with worse graphics and no food court. I held. Diamond hands. That's what we said. "Diamond hands." It means refusing to sell while your investment loses 94% of its value. We turned financial paralysis into a personality trait. A guy in my Discord paid $2.4 million for a 618-parcel estate in Decentraland. Prime district. High foot traffic. I asked him what "foot traffic" meant when the platform had 38 daily active users. He said I didn't understand the technology. I didn't. I still bought more. We had a DAO. A decentralized autonomous organization. That means we voted on decisions. There were nine of us. Three never showed up. Two voted on everything without reading it. The other four were me and my alts. We voted to "acquire strategic parcels." The vote passed unanimously. I voted four times. My portfolio peaked at $1.2 million. I told everyone. I made a spreadsheet. I projected 40x returns by 2025. I made a pitch deck. The pitch deck had a slide that said "WE ARE BUILDING THE DIGITAL ECONOMY." The slide had a rocket emoji. That was my entire financial model. In 2023 I bought a Bored Ape for $189,000. It's worth $14,000 now. I don't talk about the Ape. I still use it as my profile picture. People ask me about it. I say "I'm long-term bullish." Long-term bullish means I can't sell it without crying in a Panera. My mom asked me what a Bored Ape was. I said "digital art on the blockchain." She asked why it cost more than her car. I said "you don't understand Web3." She said "I understand you live in a studio apartment." She's not in my Discord. Justin Bieber bought one for $1.3 million. It's worth about $90,000 now. I felt better about mine after I heard that. That's community. WAGMI. We're All Gonna Make It. We said that every day. In the group chat. While the floor dropped. While the volume dried up. While 95% of all NFT collections went to zero. We're all gonna make it. None of us made it. But we said it with conviction and a laser-eye profile picture. That counts for something. It doesn't. But we said it did. That's decentralized consensus. Meta spent $84 billion on the metaverse. I need to say that again. $84 billion. More than the GDP of Luxembourg. More than the GDP of Iceland, Luxembourg, and Malta combined. They spent it on a platform where the avatars had no legs, the graphics looked like a 2006 Wii game, and the peak user count was lower than the lunch rush at a Chipotle in Des Moines. They just pulled Horizon Worlds from VR headsets. It lives on as a mobile app. My beachfront villa is now a mobile app. Location, location, location. Zuckerberg renamed the entire company for this. Facebook became Meta. A $900 billion company changed its legal name because the CEO watched Ready Player One and said "I want that." Reality Labs lost $10 billion in 2021. $14 billion in 2022. $16 billion in 2023. $18 billion in 2024. $19 billion in 2025. That's not a strategy. That's a speedrun. They laid off 1,500 Reality Labs employees this year. Shut down three VR studios. Killed Supernatural. Put the entire VR social vision in a casket and said "we're pivoting to AI and wearables." The pivot took four years and $84 billion. I pivoted too. I'm an AI real estate investor now. I bought a virtual plot in an AI-generated world that doesn't exist yet. The founder said it was "the intersection of spatial computing and large language models." I don't know what that means. I gave him $40,000. He has a whitepaper. It's 47 pages. I read the title and the tokenomics section. The tokenomics section is a pie chart. I love pie charts. They make everything look like a plan. The project has a roadmap. Q1: "Build community." Q2: "Launch beta." Q3: "Scale ecosystem." Q4 is blank. Q4 is always blank. That's where the exit scam goes. My accountant asked me to value my metaverse portfolio for tax purposes. I said $1.2 million. He said "current market value." I said $6,400. He stared at me for eleven seconds. I know because I counted. He asked if I had any other investments. I showed him my NFTs. He stared for longer. I told him they were "cultural artifacts with long-term provenance." He asked if I'd considered a 401k. I told him a 401k was "legacy finance." He told me to leave his office. The metaverse is dead. I don't accept that. I am a digital land baron. I own eleven properties across four platforms. I have a beachfront villa in a mobile app, a plot next to an empty Gucci store, and a cartoon monkey that cost me more than my actual car. Location, location, location. The location is nowhere. But I'm early. I'm always early. That's the same as being wrong except you get to say it with confidence.

@Cloudflare Thanks! Please stop this madness

Italy’s "Piracy Shield" forces providers to block content in under 30 minutes without judicial oversight, which leads to overblocking (taking down legitimate websites alongside infringing ones). We're appealing a €14M fine to protect the Internet from automated censorship and ensure infrastructure providers aren't forced to overblock. cfl.re/4cMh0WA

@steveyun_ton Wow

Founders on TON should wake up from a delusion. You are not Telegram. You are just a mini app on Telegram. Unless you offer a uniquely compelling service, no one will use you. Be honest. If you are just a copy paste, go buy a lottery instead.

@sudoingX Rtx 5090 laptop (blackwell, 8gb)

drop your GPU below. i'll tell you exactly what model and config to run on it. here's what i've tested and verified on real hardware: RTX 3060 12GB - Qwen 3.5 9B Q4 - 50 tok/s - 128K context RTX 3090 24GB - Qwen 3.5 27B Q4 - 35 tok/s - 300K context RTX 3090 24GB - Qwen 3.5 35B MoE Q4 - 112 tok/s - 262K context 2x RTX 3090 - Qwen3-Coder 80B Q4 - 46 tok/s - full VRAM all running llama.cpp with flash attention. every number is real. every config is tested. if your card isn't on this list drop it below and i'll tell you what fits.

Nope, tailscale is not recognized as secure binding, if you want to use tailscale you need to configure ssl on the gateway. Best way is to use loopback binding for both gateway and node (ws://, http:// no ssl required) and create an ssh tunnel between gateway and node over tailscale.

@croll83 @openclaw Thanks. I presume tailscale solves the security concerns?

OpenClaw 2026.3.13 🦞 👁️ live Chrome session attach — real logins, one toggle, zero extensions 📱 android redesigned & down to 7MB, iOS gets welcome pager 🐳 docker timezone override 🪟 windows gateway tweaks the lobster sees all now github.com/openclaw/openc…

You have to setup your machine (or another) as an openclaw node (install openclaw package and configure gateway mode: remote). The vps openclaw will basically use the openclaw process on your local node to use the non headless browser. This config is not easy from a sec standpoint: the node should bind localhost. The gateway should not bind public ip address, but localhost. To let them talk securely you need a tunnel between the 2 machines

@openclaw Dumb question, will this work on headless setups? Or tunnel to Chrome on a different device?