@sujee.dev · Mythryl

sujee.dev

@sujee.dev

57 followers 67 following130 posts

sujee.dev

@sujee.dev

Friday fun! 🎨🤖 I built Doodle Derby - where 3 vision models compete to recognize doodles as they're being drawn in real time! @MiniMax_AI M3.1-Flash @deepseek_ai V4.1-Flash-1017 @NVIDIAAI Nemotron-3.5-Super-VL All running on Nebius TF Who won? And what were the models thinking? 👀 Watch 👇

sujee.dev

@sujee.dev

And the results are in! 🏆 It's a tie! 🎉 The second image is my favorite - a peek into what the models were guessing along the way. Some of those guesses are hilarious 😂

sujee.dev

@sujee.dev

🎨 Your turn! BYOK (Bring Your Own Key) and play Doodle Derby: sujee.github.io/doodle-derby/ 🔑 Keyless version coming soon! 💻 Source: github.com/sujee/doodle... Give it a try! Feedback and suggestions welcome! 🙌

sujee.github.ioDoodle Derby

sujee.dev

@sujee.dev

What actually happens when you send a request to an LLM? Client → Gateway → tokenizer → router → queue → prefill → decode → streamed back to you, one token at a time. I'm tracing the whole trip tomorrow at Hacktoberfest 2026. ⚡ Thu, Oct 22 online @ 10am PT / 1pm ET dev.to/events/hackt...

sujee.dev

@sujee.dev

Friday fun 🎉 The Nemotron family (Super, Lightning, Ultra) went to the carnival… and raced horses 🎪🏇 One of them is a total sprinter. Guess who? 👀 Race your own models 👉 sujee.github.io/llm-derby

sujee.dev

@sujee.dev

My talk with Dylan at AI Engineer World’s Fair SF is now live! “Optimizing Open Models for Production” covered: - KV caching - Cache-aware routing - Speculative decoding - Disaggregated prefill + decoding - Quantization 🎥 Talk: www.youtube.com/watch?v=TRe1... 📊 Slides: sujee.dev/events/2026-...

youtube.comWhat Makes Open Models Fast in Production — Sujee Maniyam, NebiusYouTube video by AI Engineer

sujee.dev

@sujee.dev

Friday fun! LLM Derby 🏇 puts models head-to-head on the same questions. Every correct answer moves a horse forward. First to 20 wins. This race: 🏇 Qwen3.8-27B 🏇 gpt-6-astra 🏇 claude-sonnet-5-5 Who won? Watch the video 👇 Try the game: sujee.github.io/llm-derby/ Just a fun game, not a benchmark 😄

sujee.dev

@sujee.dev

Watch: youtu.be/Hk8Dn0qfMT4 Play: sujee.github.io/llm-derby/ code: github.com/sujee/llm-de...

youtu.beLLM Derby: Qwen3.8-27B vs gpt-6-astra vs claude-sonnet-5-5 trivia raceYouTube video by sujee_dev

sujee.dev

@sujee.dev

1/ A few of you asked how I made the Linux history video 🐧 One prompt in Claude Code (Opus 5.5): "Create visualization tracing Linux history. A 2 min video, with cool music." No storyboard. No extra skills, MCP servers or API keys. www.youtube.com/watch?v=lkOx...

youtube.comLinux Turns 35: From "Just a Hobby" to Running the WorldYouTube video by sujee_dev

sujee.dev

@sujee.dev

2/ ~7 min later: a 2-minute video with a soundtrack generated in Python. ~80% there. The rest was notes like "0:30: keep this sequence longer". Prompt, code, and how to remix it: github.com/sujee/visual... Anyone else making motion graphics this way?

github.comvisual-stories/projects/linux-history at main · sujee/visual-storiesOpen-source, reproducible visual explainers for AI and software concepts. Watch them. Reproduce them. Remix them. - sujee/visual-stories

sujee.dev reposted

All Things Open

@allthingsopen.bsky.social

We're thrilled to have Shivay Lamba, Lead ML Engineer for Qualcomm, presenting "AI Sandboxes: Running Coding Agents Safely in Production-Grade Environments" at #AllThingsOpen! 2026.allthingsopen.org/sessions/ai-...

sujee.dev

@sujee.dev

35 years ago today, Linux 0.02 was released 🐧 I've been using Linux for ~31 of those years. 1995: recompiling kernels to get new hardware working. 2026: Linux Mint + NVIDIA GPU, running LLMs locally. I made this little visual history of Linux 👇 What’s your favorite Linux memory?

sujee.dev

@sujee.dev

Found this one in the archives :-) Me holding a Red Hat Linux 5.1 CD, circa 1998. Check out those CRTs! I wrote a little more about my 31-year journey with Linux: sujee.dev/post/linux-3... Also 4K video : youtu.be/lkOxEeqwMBo

sujee.dev

@sujee.dev

Most AI models write. Jev decides. A 2-minute explainer for devs. Made with Claude Code + Opus 5.5. Open source and fully reproducible. Remix it: github.com/sujee/visual... #AI #LLM #ExplainerVideo #Jev

sujee.dev

@sujee.dev

Also on YouTube (4K, with chapters): youtu.be/RQ6f6z-dPWA More explainers: www.youtube.com/playlist?lis...

youtu.beJev vs. LLMs: a model that decides instead of writing | Explainer VideoYouTube video by sujee_dev

sujee.dev

@sujee.dev

Quick demo showing the vision capabilities of Qwen 3.8-27B 👇 I gave it a crinkled, scanned-looking document containing PII. It extracted the text, analyzed the document, and identified the PII ✅ This little model is a beast 😄 And it was ticking along at 272 tok/sec on @nebiustf 🚀

sujee.dev

@sujee.dev

Here’s the crinkled/scanned-looking document image (generated!) I fed into the model 👇 Not exactly a clean input - which makes the result even more fun.

sujee.dev

@sujee.dev

Friday funsies 🐍 LLM Snake Arena: GPT-6-Astra vs Claude Opus 5.5 Does it tell us something about the models? Maybe. Fun to watch? Most definitely 😀

sujee.dev

@sujee.dev

Try it for yourself 🐍 sujee.github.io/llm-snake-ar... Go burn some tokens :-) Code: github.com/sujee/llm-sn...

sujee.dev

@sujee.dev

16 Sept

Want to try open models for coding? Keep your coding agent - Claude Code, Codex, OpenCode, Cline - and just swap the model underneath. In this short talk: * open vs closed models * why the coding harness matters * what coding tasks cost * quick setup demo Watch: www.youtube.com/watch?v=gjNc...

youtube.comCoding with Open Models: Claude Code + OpenCode with Kimi, GLM & DeepSeekYouTube video by sujee_dev

sujee.dev

@sujee.dev

16 Sept

Setup + instructions: github.com/sujee/worksh... Want to try open models? Get credits: dev.nebius.com/builders?utm...

github.comworkshops/coding-with-open-models/README.md at main · sujee/workshopsHands-on workshops on AI engineering, open models, coding agents, LLM inference, developer tools, and physical AI. - sujee/workshops

sujee.dev

@sujee.dev

12 Sept

Artificial Analysis just updated its Intelligence Index to v4.3 with a tougher benchmark suite. Scores dropped quite a bit. Top proprietary models now: - Claude Fable 5.1: 53 = 66 → 53 = ▼ 13 - GPT-6 Astra: 53 = 61 → 53 = ▼ 8

sujee.dev

@sujee.dev

12 Sept

Top open models now: 1. GLM-5.3: 45 2. Kimi K3: 44 (60 → 44 = ▼ 16) 3. GLM-5.3-Flash: 42 (57 → 42 = ▼ 15 ) Big score shifts, but the important thing is the relative ranking under the new benchmark suite.

sujee.dev

@sujee.dev

7 Sept

Went to buy a 4TB portable SSDs and almost fell off my chair looking at the prices 😳 Damn. See screenshots 👇 @demian_ai has been writing about the memory/storage crunch driven by the AI boom. It is finally sinking in for me :-) x.com/demian_ai/st... x.com/demian_ai/st... x.com/demian_ai/st...

sujee.dev

@sujee.dev

7 Sept

Fine-tuning, distillation, and quantization are easy to blur together. Here is an explainer video with LEGO® bricks 🧱 🟢 Fine-tune = specialize 🟠 Distill = teach a student 🟡 Quantize = use lower precision More detail: sujee.dev/post/fine-tu... drop your suggestions in comments

sujee.dev

@sujee.dev

29 Aug

1️⃣ First graphic: GLM-5.3-Flash (formerly Ox Alpha!) scores 57 on the Artificial Analysis Intelligence Index, while Kimi K3 scores 60. But GLM-5.3-Flash is: much cheaper, much smaller Try it for yourself : tokenfactory.nebius.com/playground?m... Model visualizer: sujee.github.io/practical-ll...

sujee.dev

@sujee.dev

29 Aug

2️⃣ Second pic: GLM-5.3-Flash running inside OpenCode. Amazing speed and performance! ⚡

sujee.dev

@sujee.dev

28 Aug

I tested 3 NVIDIA Nemotron models with the same sales-analysis agent. All scored 100%. Lightning done in 7s for $0.0019 - about 7x cheaper than Super and 16x cheaper than Ultra. TLDR; For focused agent workflows, smaller may be enough. www.youtube.com/watch?v=Ycme... github.com/nebius/token...

sujee.dev

@sujee.dev

20 Aug

🚀 Token Factory Office Hour #4 Join us live with special guest Chris Alexiuk from NVIDIA talking Nemotron models. Also * What’s new in @nebiustf * Builder demos * Live Q&A * Credits 🎁 Like to show what you built with Token Factory? DM me. 📅 Aug 26, 9 AM PDT Register: nebius.com/events/webin...

nebius.comToken Factory office hours: Nemotron modelsBuild and scale faster on the purpose-built AI cloud, engineered from silicon to API.

sujee.dev

@sujee.dev

15 Aug

I am doing quick hands-on workshop on ClawCamp SF Summer Summit tomorrow! 🦞 “Coding with Open Models + Coding Agents” Want to try Kimi K3, GLM 5.2 and other open models? Bring your laptop 💻 We’ll provide Token Factory credits so you can try them with your coding agent. luma.com/clawcamp-sf-...

luma.com🦞 ClawCamp SF Summer Summit 🦞 · LumaNEW VENUE: Ecosystem Coworking 540 Howard St 2nd floor, San Francisco, CA 94105 Join us to Up-Level Your Agentic AI Superpowers. All Levels Welcome. This…

sujee.dev

@sujee.dev

7 Aug

Kimi K3 was one of the most anticipated open models. Join the Kimi + Nebius Token Factory teams on Aug 12: 🏗️ K3 architecture + capabilities ⚡ Inference engineering + optimizations 🎁 Token Factory credits to try K3 yourself 9 AM PT / 12 PM ET / 6 PM CET 👉 nebius.com/events/webin...

sujee.dev

@sujee.dev

31 July

Friday fun: Here are some cool @kimi-moonshot-x.bsky.social K3 demos I’ve seen so far 👇 Charlie and the Token Factory by @demian_ai 🎮 datamon.vercel.app Subway Surfers x.com/Arindam_1729... Car crushing and physics simulation x.com/atomic_chat_... fish tank x.com/UnslothAI/st...

datamon.vercel.appCharlie & the Token FactoryA full browser RPG set inside the Nebius Token Factory.

sujee.dev

@sujee.dev

31 July

3D map of New York City x.com/kirillk_web3... And a browser-playable Quake game from me 🎮 x.com/sujee_dev/st... I’m collecting more fun Kimi K3 demos here: github.com/sujee/practi... Know of another cool demo? Drop it in the comments!

x.comKirill (@kirillk_web3) on XKimi K3 is dangerous. I rebuilt Google Maps 3D mode in 1.5 hours. an interactive 3D map — New York City in full detail, and the entire globe on top of it. two prompts. that's the whole build. > ...

sujee.dev

@sujee.dev

28 July

Trying out @kimi-moonshot-x.bsky.social K3 (running on Nebius Token Factory ) for coding with @opencode.bsky.social . The integration was incredibly smooth - easy to configure and start building. Here’s a quick 30-second walkthrough to get you up and running. 👇 start: tokenfactory.nebius.com

sujee.dev

@sujee.dev

29 July

Join the Nebius AI Builder Program and get $25 in free Token Factory credits to start building: dev.nebius.com/builders?utm...

dev.nebius.comNebius Builder ProgramBuild and scale faster on the purpose-built AI cloud, engineered from silicon to API.

sujee.dev

@sujee.dev

23 July

Nebius Token Factory @nebiustf Office Hour #3 Recap: We covered new models and features, inference optimizations, fine-tuning, the Builder and Fellow programs, plus live Q&A. recording + slides in comment 👇 See you at the August office hour: nebius.com/events/webin...

sujee.dev

@sujee.dev

23 July

🎬 Replay: nebius.com/events/webin... 📑 Slides: email.nebius.com/hubfs/webina...

nebius.comToken Factory office hours: Build and ship with open modelsBuild and scale faster on the purpose-built AI cloud, engineered from silicon to API.

sujee.dev

@sujee.dev

22 July

Building with open models? Join us for Nebius Token Factory Office Hour - a live, developer-focused session featuring demos and Q&A 🗓️ July 22, 2026 ⏰ 9 AM PT / 12 PM ET You’ll also receive Token Factory credits, so you can start building right away. See you there! nebius.com/events/webin...

sujee.dev

@sujee.dev

2 July

Nebius booths are never boring. Meet NEBU - our Nebius pup 🐶 - making a new friend 🐉 Stop by the Nebius booth at AI Engineer World Fair. You can play with the robots... and chat with the humans, too. 😄

sujee.dev

@sujee.dev

1 July

Come by the Nebius booth at AI Engineer conf! We’ve got fun demos: LLMs playing Snake 🐍, LLMs playing soccer/fútbol ⚽️ Demo below 👇 We’ll also show you our open-model coding stack - GLM 5.2 and Kimi K2.6 are our current faves. Plus, we’ll set you up with Token Factory credits 🎁

sujee.dev

@sujee.dev

23 June

I’m running a workshop at ClawCamp Campout @ Dual Tech Summit! We’ll add real-time intelligence to OpenClaw using Nebius Token Factory + Tavily 🗓️ Wed, June 24, 2026 📍 War Memorial Veterans Building, SF 👉 clawcamp.us/event/260624... 🎁 Free Token Factory + Tavily credits . See you there!

sujee.dev

@sujee.dev

19 Feb

The AgentBeats–AgentX Competition Phase 1 winners have been announced - and the projects are fantastic. Huge congratulations to all the teams involved! 👏 Read the highlights here: berkeleyrdi.substack.com/i/188179396/... Now, on to Phase 2 : rdi.berkeley.edu/agentx-agent...

sujee.dev

@sujee.dev

18 Feb

(another car wash post) Q: I want to wash my car. The car wash is 50 meters away. Should I walk or drive? Kimi K2.5: (thought 52 secs) : Drive the 50 meters, but maybe take the long way around the block to let the engine warm up slightly ...

sujee.dev

@sujee.dev

18 Feb

GLM 4.7 (thought 42 secs): If your car has a manual transmission (stick shift) and the road is flat, you might choose to **push** the car to save fuel and avoid the wear of starting the engine just for a 50-meter drive ... (see image for full answer) sometimes the models do overthink :-)

sujee.dev

@sujee.dev

13 Feb

Kimi 2.5 from @kimi-moonshot-x.bsky.social is now live on Nebius Token Factory. A major step up from Kimi 2, this os multimodal model is strong in agentic workflows and coding tasks - and it’s stacking up well against SOTA proprietary models. Check out the official benchmarks and the pelican 🪿🚲

sujee.dev

@sujee.dev

13 Feb

Try it out: tokenfactory.nebius.com/playground?m...

tokenfactory.nebius.comNebius Token FactoryTransform your business with AI on Nebius Token Factory. Premium open-source LLMs and image models, simple API integration, and infrastructure that scales automatically with your growth.

sujee.dev

@sujee.dev

15 Jan

MiniMax M2.1 from MiniMax is now live on Nebius Token Factory. strong showing for coding, agentic workflows It’s also ranked among the **top open models** by Artificial Analysis And yes… it passes my favorite vibe benchmark too: “pelican riding a bicycle” 🪿🚲 - look at that beak 😄

sujee.dev

@sujee.dev

13 Jan

qwen3-next-80B-A3B-Thinking is now available on Nebius Token Factory. “Fast but verbose” per Artificial Analysis: artificialanalysis.ai/models/qwen3... My fav vibe benchmark: 'pelican riding a bicycle' 🪿🚲 Try it 👉 tokenfactory.nebius.com/playground?m... #OpenModels #LLM #PelicanRidingBicycle

sujee.dev

@sujee.dev

8 Jan

Installed @linuxmint.bsky.social on desktop & laptop Overall - smooth install - great collection of OS software Desktop: - detected dual monitors and NVIDIA GPU Laptop: - every thing works including suspend / wake A pic of my two fav mints :) Install notes: sujee.dev/post/another... #linux

sujee.dev

@sujee.dev

3 Dec 2025

🚀 Token Factory Builder Hour #1 Dec 9 @ 9am PT / 12pm ET / 5pm UTC / 6pm CET Hang out, share ideas, see demos, learn what’s new in Token Factory. Can’t make it? We’ll send out the recording + notes. Join: nebius.com/events/webin... Discord: [https://discord.gg/2BUfPAu5uH](discord.gg/2BUfPAu5uH)

nebius.comBuilder hour: Token FactoryDiscover the most efficient way to build, tune and run your AI models and applications on top-notch NVIDIA® GPUs.