Sujee Maniyam (@sujee_dev)

@sujee_devOct 29

👋 I build with AI in the open, explore open models, and share what I learn. Here’s what I’m working on: LLM benchmarks & evals ⚡ ZebraBench: benchmark LLM inference right in your browser sujee.github.io/zebrabench/ 🧪 Practical LLM Evals: tools and examples for evaluating LLMs github.com/sujee/practical-l… Playful AI demos 🎨 Doodle Derby: you draw, vision models race to guess sujee.github.io/doodle-derby… 🏇 LLM Derby: models race to answer trivia sujee.github.io/llm-derby/ Visual explainers 🎬 Visual Stories: animated explanations of AI and software concepts github.com/sujee/visual-stor… Follow along as I build, experiment, and share what I learn. 🛠️ More projects: sujee.dev/projects/ 🎤 Talks & workshops: sujee.dev/speaking/

220

Sujee Maniyam · Oct 30, 2026 · 9:08 AM UTC

@sujee_devOct 30

This one's for my Tamil-speaking friends! Trying out Agaram 2.5B (அகரம்) - a Tamil-English language model trained from scratch by @MuraiLabsHQ & @Mechramc. Running locally on my Mac with Ollama! 🦙 First, some casual chatting. English: >>> how are you? > I'm doing well, thank you for asking! Tanglish (Tamil written in English letters): >>> eppadi suham? > I'm doing great! Tanglish, requesting a Tamil response: >>> eppadi suham? answer in tamil > நான் நல்லா இருக்கேன். உங்களுக்கு என்ன உதவி வேணும்னு சொல்லுங்க, நான் பண்றேன். Even got it to write a poem about cats! 🐈 >>> write me a poem about cats > நிச்சயமாக! இதோ ஒரு குட்டி கவிதை: > வீட்டின் காவலன், > வெயிலில் உழைக்கும் நண்பன், > அன்பான நண்பன், > என் செல்லக்குட்டி! Pretty cool for a 2.5B model! But then I got curious: How well does it understand different Tamil dialects, particularly Sri Lankan Tamil vs. Tamil Nadu Tamil? One question led to another, and before I knew it, I was down a rabbit hole testing Tamil dialects! 🐇 Here's what I found. 🧵👇

175

Sujee Maniyam · Oct 30, 2026 · 9:10 AM UTC

@sujee_devOct 30

So I turned my curiosity into a little experiment! 🇱🇰🇮🇳 I put together 20 everyday phrases, each written in Sri Lankan Tamil and Tamil Nadu Tamil, with the same meaning. Then I asked Agaram 2.5B to translate all 40 sentences into English. Results: 🇱🇰 Sri Lankan Tamil: 12/20 ✅ (60%) 🇮🇳 Tamil Nadu Tamil: 17/20 ✅ (85%) see screenshot Some interesting regional vocabulary differences: - வடிவா vs. அழகா (beautiful) - கதைக்க vs. பேச (to talk) - இருங்கோ vs. உக்காருங்க (sit down) The model recognized quite a few Sri Lankan Tamil expressions! My guess is that training data plays a role here. Tamil Nadu Tamil likely has more representation in the training corpus than Sri Lankan Tamil. This is just a small, informal experiment, not a comprehensive benchmark. But it was fascinating to see how regional language variations affect model understanding! Huge kudos to @Mechramc and @MuraiLabsHQ for building and open-sourcing a Tamil-first model! 👏 It's exciting to see more AI development focused on languages like Tamil. I've shared my experiment, including the test data, Python code, and full results. Feel free to reproduce it, extend it, or share your feedback! 😁 github.com/sujee/ai-playgrou…

127

Sujee Maniyam · Oct 30, 2026 · 9:14 PM UTC

@sujee_dev17h

@Mechramc Curious about what's next for Agaram! 😁 - More training, perhaps with broader coverage of regional Tamil dialects? 🙂 - System prompt support to customize the model's personality and behavior? Excited to see where this project goes!

Sujee Maniyam · Oct 30, 2026 · 8:18 PM UTC

@sujee_dev18h

What does it take to run Kimi K3, a massive 2.8-trillion-parameter model, for inference? Enjoyed this whiteboard walkthrough by @seanlie and @MilksandMatcha from @Cerebras. Love the warehouse/lab setting too :-) Here’s my takeaway from their Kimi K3 example. K3 has 93 layers: • 69 KDA (Kimi Delta Attention) layers using linear attention • 24 Gated MLA (Multi-head Latent Attention) layers using full attention with a compressed KV cache The memory distinction matters: KDA maintains fixed-size state per request. MLA keeps a KV cache that grows with context length. Now the math: 2.8T parameters at 4 bits each ≈ 1.4 TB of raw weights. Following Cerebras’ example, let’s round that to 3T parameters: • 3T × 4 bits ÷ 8 ≈ 1.5 TB • 1.5 TB ÷ 93 layers ≈ 16 GB/layer on average Each wafer has 44 GB of on-chip SRAM. Cerebras’ illustrative mapping is: • Two linear-attention layers (16 GB each) per wafer: ~32 GB of weights, leaving ~12 GB • One full-attention layer per wafer: ~16 GB of weights, leaving ~28 GB for KV cache and other state That gives: • 69 linear-attention layers ÷ 2, rounded up = 35 wafers • 24 full-attention layers ÷ 1 = 24 wafers • Total: 59 wafers This example assumes 4-bit weights, an 8-bit KV cache, and 256K average request lengths. The neat part: weights stay on-chip while activations move through the wafer pipeline. Larger models? Like 5T or 10T? :-) Extend the pipeline with more wafers. The exact count depends on the model architecture and memory requirements. Highly recommend watching the full video!

Sean Lie

@seanlieOct 29

How we deploy multi-trillion parameter models at scale Today, let's break down the math for how @cerebras: 1. Fits the full model into on-chip memory 2. Maps the model across multiple chips 3. Handles a high # of concurrent requests

262

Sujee Maniyam · Oct 30, 2026 · 8:19 PM UTC

@sujee_dev18h

Curious where the KV cache comes into play? In this video, I follow a single request from client to chip and back, including how the KV cache fits into inference. nitter.cf/sujee_dev/status/21152…

What actually happens when you call an LLM API? 🤔 I've been working on this visual explainer for a while, and I'm excited to finally share it! 🎬 Follow a single request from client → gateway → tokenizer → router → queue → GPU (prefill + decode) → streamed response. Along the way, explore optimizations like: KV caching / Prompt caching Cache-aware routing Continuous batching Disaggregated prefill/decode Speculative decoding I recently gave a talk on this at Hacktoberfest 2026, but here's the entire journey in just 5 minutes! Watch it and let me know what you think! 👇 Full talk and repo in the reply.

Ramchand Kumaresan · Oct 28, 2026 · 12:52 PM UTC

Sujee Maniyam retweeted

Ramchand Kumaresan

@MechramcOct 28

Agaram 2.5B is out. It's a Tamil-first language model we trained from scratch at @MuraiLabsHQ , our small non-profit lab. As far as we know, it's the first Tamil model trained from scratch that can call tools. It chats in Tamil, Tanglish and English, translates, and can call tools. We're releasing the base model, chat model and GGUF files under Apache-2.0. You can download it and run it offline on your own laptop. The first chat model had a four-day public preview. Seeing where it fell short with real users helped us decide what to work on. We rebuilt the fine-tuning around what we saw, using 222,662 records. Thank you to everyone who spent time with that early version. The video walks through the model. You can try it here: hf.co/spaces/mechramc/agaram… நன்றி 🙏

113

421

2,277

91,011

Sujee Maniyam · Oct 29, 2026 · 10:27 PM UTC

@sujee_devOct 29

Welcome to the 1T parameter club, @MistralAI Le Chonk! 🐈

Mistral AI

@MistralAIOct 29

Today, we release the weights of a new and improved Le Chonk, advancing the open frontier for agentic coding and cybersecurity. 💻 Built to build: Le Chonk ranks #1 on @cognition’s FrontierCode among open models, and sets a new benchmark for Pareto efficiency on @artificialanalysis. 🛡 Built to defend: State-of-the-art in cyberdefense, put through the paces by the world's leading cyber vendors, including @CrowdStrike, @cloudflare, @rubrik, @zscaler, and others. Built for sovereignty: Make it yours. Download the weights on @huggingface , customize it with Mistral Forge, and deploy it wherever you need it. Read more from our team and customers 👇:

188

Sujee Maniyam · Oct 28, 2026 · 6:22 AM UTC

@sujee_devOct 28

Thanks for the shoutout @ChrisPirillo 😊

Chris Pirillo

@ChrisPirilloOct 28

Vibe Coding & AI News to Use: October 27, 2026… A visual explainer traces an LLM API request to the GPU and back. @sujee_dev github.com/sujee/visual-stor… Better Backlog filters GitHub issues with decision models. github.com/BerriAI/better-ba… @AnthropicAI shares early progress on drug development with Claude. anthropic.com/research/early… OpenAI's Codex CLI reaches rust-v0.165.0. github.com/openai/codex/rele… Gemini CLI v0.67.0-preview.0 rolls out a new preview build. @googleaidevs github.com/google-gemini/gem… @hexanalytics raises $100M. hex.tech/blog/hex-new-era/ CleanCode puts Claude Code and Codex on one shared canvas. github.com/chen-985211/clean… Agent Skills collects 51 SKILL.md workflows for coding agents. github.com/milad-ahmd/agent-… OpenBot 0.52.1 refreshes a desktop workspace for persistent AI teammates. github.com/nightly-labs/open… @discord will soon surface public server messages through Google. discord.com/blog/making-cont… Wenlan is an LLM wiki that flags what it can't back up. github.com/wenlan-app/wenlan Ghostty 1.4 is out. ghostty.org/docs/install/rel… Genie 4 opens a new generation of DeepMind world simulation. deepmind.google/blog/genie-4… OpenHands 1.34.0 updates the open-source coding agent. github.com/All-Hands-AI/Open… @jetbrains ThinkRail ships a new v0.3.0 nightly of its vibe-coding IDE. github.com/JetBrains/thinkra… Wikipedia's cleanup project tackles AI-generated writing. @ieeespectrum spectrum.ieee.org/wikipedia-… Cindy v0.1.118-beta iterates on an out-of-the-box open-source agent. github.com/makecindy/cindy/r… PulseDev: click a UI bug on localhost and your agent fixes it. pulsedev.app Revix AI reviews code against your repo's own rules. @userevix revix.ai Try this image style: Take the attached photo and redraw it as Needle-felted wool diorama: eyes two round buttons of dense black wool framed by stitched ring rims, nose a softly pinched wool wedge, mouth one strand of roving couched in a closed gentle smile, hair a close cap of felted charcoal fibres with a soft halo of loose fleece, and hands and body shaped from firmly needled oat-coloured wool that keeps its visible felting-needle pockmarks. Palette: oat fleece, madder red, slate blue, charcoal wool. Waist-up three-quarter view of Chris at a wide shared table, needle-felting the last seam of a small madder-red helper robot while its slate-blue felt mate reaches across a single stretched canvas lying flat between them, loose wisps of raw fleece and a barbed needle resting at the table edge, warm side light raking the fuzzy wool halo. Keep the person recognizable. Square, no text. Meet Seattle-area builders at Builder Day and Builder Night. See the calendar for upcoming events. luma.com/ctrlaltcreatelive ctrlaltcreate.live/

ALT Chris Pirillo felting helper robots that share a single canvas. Needle-felted wool diorama illustration.

182

Sujee Maniyam · Oct 27, 2026 · 11:17 PM UTC

Want to try @Kimi_Moonshot K3.1 Pro? Here are a few ways to access it: 1 - OpenCode Go Starts at $10/month. An affordable way to get started. opencode.ai/go 2- Kimi CLI + Kimi API $2/M input tokens, $10/M output, $0.20/M cached input platform.kimi.ai/docs/pricin… 3 - OpenRouter Same API pricing. openrouter.ai/moonshotai/kim… 4 - Find other providers Compare options on models.dev: models.dev/models/moonshotai… I've tried both Kimi CLI + Kimi API and OpenCode + OpenCode Go. My experience so far: - Kimi CLI has been the most reliable for me. - OpenCode Go is a budget-friendly way to get started, but it's been a little laggy, and you could hit usage limits during intensive coding tasks. How are you using K3.1 Pro? What's your experience been like? (Note: K3.1 Pro weights aren't public yet. @Kimi_Moonshot says they'll be released by Nov 4, 2026.)

OpenCode Go | Low cost coding models for everyone

Go costs $10/month, with generous usage limits and reliable access to leading coding models. Go Plus costs $40/month with higher limits.

opencode.ai

162

Sujee Maniyam · Oct 27, 2026 · 10:38 PM UTC

1,514

more replies

Sujee Maniyam · Oct 27, 2026 · 10:40 PM UTC

🛠️ Behind the scenes: This video is part of my Visual Stories series, where I use AI-assisted animation and storytelling to make complex technical concepts easier to understand. The project lives on GitHub, along with the source code and production details. Check it out & remix & share: github.com/sujee/visual-stor…

Sujee Maniyam · Oct 27, 2026 · 10:46 PM UTC

@hacktoberfest ☝️

Sujee Maniyam · Oct 23, 2026 · 8:55 PM UTC

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 @nebiustf 🚀 Who won? And what were the models thinking? 👀 Watch 👇 (Inspired by Google's Quick, Draw!)

345

Sujee Maniyam · Oct 23, 2026 · 8:56 PM UTC

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 Maniyam · Oct 23, 2026 · 8:58 PM UTC

🎨 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-derb… Give it a try! Feedback and suggestions welcome! 🙌

Sujee Maniyam · Oct 23, 2026 · 7:23 PM UTC