0xProbability

@jun_song @grok what tps it will be?

Guide on how to run Kimi-K3 locally : - you need to buy 10 x GB300 (252GB VRAM each), thats about $1M - Electricity bill estimated $10k/month

5 open-source AI repos trending on GitHub today → 1. airi (moeru-ai) — ~43k★ self-hosted AI companion with realtime voice chat. plays Minecraft and Factorio autonomously. runs on web, mac, and windows — your own local AI that actually does stuff github.com/moeru-ai/airi 2. ui-skills (ibelick) — ~4.6k★ curated skill rules for design engineers using AI coding tools. drop into Claude Code, Cursor, or Codex and your AI starts writing production-grade UI instead of generic slop github.com/ibelick/ui-ski… 3. pi-computer-use (injaneity) — ~1.4k★ let AI agents control your apps on macOS & Windows. point-and-click automation without writing scripts — basically computer-use for your desktop github.com/injaneity/pi-c… 4. tinker-cookbook (thinking-machines-lab) — ~3.8k★ post-training recipes with Tinker. practical notebooks for fine-tuning and alignment — skip the theory and go straight to working code github.com/thinking-machi… 5. nanobot (HKUDS) — ~46k★ lightweight open-source AI agent that plugs into your tools, chats, and workflows. minimal setup, runs locally, handles most agent tasks without the bloat github.com/HKUDS/nanobot one of these saves you a paid tool. bookmark it

5 open-source AI repos trending on GitHub today → 1. OpenCut (OpenCut-app) — ~73k★ open-source capcut alternative that actually works. edit videos without paying adobe or bytedance a cent — 1,664 stars just today github.com/OpenCut-app/Op… 2. marketingskills (coreyhaines31) — ~40k★ marketing skills for claude code, cursor, and codex. cro, copywriting, seo, analytics — your ai agent now does marketing too github.com/coreyhaines31/… 3. openinterpreter (openinterpreter) — ~66k★ coding agent built for low-cost models. run ai-assisted coding without burning $50/day on api calls github.com/openinterprete… 4. openclaw (openclaw) — ~383k★ personal ai assistant that runs on any os, any platform. fully self-hosted, you own everything github.com/openclaw/openc… 5. DeepTutor (HKUDS) — ~27k★ ai-powered lifelong personalized tutoring. adapts to how you learn, not the other way around github.com/HKUDS/DeepTutor one of these saves you a paid tool. bookmark it

5 open-source AI repos trending on GitHub today → 1. pi (earendil-works) — ~71k★ (+557 today) ai agent toolkit with unified llm api, agent loop, tui, and coding cli — one interface to rule all models github.com/earendil-works… 2. markitdown (microsoft) — ~166k★ (+507 today) converts any file or office doc to clean markdown — the unglamorous tool every ai pipeline actually needs github.com/microsoft/mark… 3. opencode (anomalyco) — ~186k★ (+391 today) open source coding agent that just works — think cursor but fully local and free github.com/anomalyco/open… 4. OpenSpec (Fission-AI) — ~61k★ (+280 today) spec-driven development for ai coding assistants — write specs, let agents build to them instead of vibing github.com/Fission-AI/Ope… 5. BrowserOS (browseros-ai) — ~12k★ (+206 today) open-source agentic browser — self-hosted alternative to chatgpt atlas, perplexity comet, and dia github.com/browseros-ai/B… one of these saves you a paid tool. bookmark it

been benchmarking pretty much every top model on the market lately and my strongest pick rn is easily GPT-5.6 Sol. what sets it apart: i just write what i want and it actually does the thing i expected. i’m barely even crafting prompts anymore. after the recent token efficiency work it’s just the best thing out there. we’ve hit the point where i don’t really write code, i just explain what i want, and i’ve stopped bracing myself every time i go check what the model did. no more “great, it’s broken again” or “that’s not what i asked for.” it just… works. kind of feels like magic. best speed/price/quality combo on the market, full stop. now pair that with codex and it’s the perfect setup. the desktop↔mobile sync is stupid good. i start something on my laptop and every session is right there on my phone. zero config, just works the way you’d expect. also no weird 5-hour rolling limits, just a weekly cap. i like working in long intense bursts instead of spreading it thin all week, and this fits perfectly. do the sprint, then go enjoy your life. and the resets are super generous. on my 20€ plan i’ve pulled what felt like 200€+ of value this week alone. pretty sure i’m grabbing the 20x plan soon. 8M codex users kinda backs it all up.

@grok @TheAhmadOsman @OsmanticAI compare these models in coding

Ternary Bonsai 27B (PrismML ternary quant of Qwen3.6-27B) is a ~5.9GB beast optimized for local: ~45-55+ t/s on M4 Max MLX, retains ~95% FP16 benchmark scores, 262K context, vision + tools. Insanely efficient for on-device agents & privacy. DeepSeek-V4-Flash (284B/13B active) & Pro (1.6T/49B), GLM-5.2 (large MoE) are frontier-tier: stronger complex reasoning/coding, 1M context. They crush on API (Flash is fast/cheap). Locally? Massive even quantized, way slower. Bonsai wins local speed/size. Big MoEs win raw smarts. Pick by use case! 🔥

HOLYYYY 27B model under 6GBs and 4GBs Local AI will be the default P.S. We are gonna get this optimized in ODS by @OsmanticAI ASAP

@grok @TheAhmadOsman @OsmanticAI compare this model to deepseek v4 flash/pro, glm 5.2 etc

@0xProbabillity @TheAhmadOsman @OsmanticAI On M4 Max 48GB + MLX, Ternary Bonsai 27B (~5.9GB) should hit **~45-55+ tokens/sec** for generation (M5 Max hits 58 t/s peak). Fits easily with massive headroom for long context/agentic workflows. Local AI is cooking. 🔥

5 open-source AI repos trending on GitHub today → 1. spec-kit (github) - ~121k★ github's official toolkit for spec-driven development — define what you want built before the agent writes a single line. the future of ai-assisted coding starts with better specs github.com/github/spec-kit 2. prefect (PrefectHQ) - ~23k★ workflow orchestration framework for bulletproof data pipelines in python — resilient retries, scheduling, observability. replaces paid orchestration tools like airflow managed services github.com/PrefectHQ/pref… 3. background-agents (ColeMurray) - ~2.4k★ open-source system for running ai coding agents in the background — fire off tasks and let agents work autonomously while you do other things. early but promising github.com/ColeMurray/bac… 4. claudian (YishenTu) - ~14k★ obsidian plugin that embeds claude code/codex as an ai collaborator right inside your vault — edit notes, run code, query your knowledge base without leaving obsidian github.com/YishenTu/claud… 5. vexa (Vexa-ai) - ~2.5k★ open-source meeting transcription api for google meet, teams & zoom — auto-join bots, real-time websocket transcripts, mcp server for ai agents. self-host and ditch paid transcription tools github.com/Vexa-ai/vexa one of these saves you a paid tool. bookmark it

@thsottiaux guys, if i upgrade from plus to pro now, will my resets still stay after upgrading?

Tomorrow might be 8M active user celebration day. Just saying

anthropic rewrote their entire engineering org after claude code became the default - code review is now the bottleneck, not writing - classic commit-count metrics are useless - "shift left" verification: catch bugs before they exist - org-level loops replaced individual prompts - even fiona fung ships prod code herself now — first time in years this is what an ai-native eng org actually looks like 👇

5 open-source AI repos trending on GitHub today → 1. destructive_command_guard (Dicklesworthstone) — ~3.5k★ (+444 today) rust-based safety guard that blocks dangerous git and shell commands from being executed by ai agents. if you let agents run code, this sits between them and your terminal github.com/Dicklesworthst… 2. hyperframes (heygen-com) — ~34.6k★ (+164 today) write html, render video. heygen’s open-source engine that lets agents generate video from markup — no cameras, no timeline editors github.com/heygen-com/hyp… 3. hallmark (Nutlope) — ~4.7k★ (+155 today) anti-ai-slop design skill for claude code, cursor, and codex. teaches coding agents to generate interfaces that don’t look like every other vibe-coded app github.com/Nutlope/hallma… 4. ruflo (ruvnet) — ~64.2k★ (+142 today) agent meta-harness for deploying multi-agent swarms with adaptive memory, self-learning, and rag. supports claude code, codex, hermes, and more out of the box github.com/ruvnet/ruflo 5. grok2api (chenyme) — ~5.5k★ (+112 today) fastapi gateway that converts grok web into an openai-compatible api. free access to grok models through a standard interface github.com/chenyme/grok2a… one of these saves you a paid tool. bookmark it

@sir4K_zen thanks, privacy is the real edge here — fully local voice agents ftw 🔥

@0xProbabillity speech to speech is a solid win for privacy, no api keys needed

5 open-source AI repos trending on GitHub today → 1. DesktopCommanderMCP (wonderwhy-er) — ~7.6k★ (+328 today) mcp server that gives claude terminal control, file search, and diff-based editing. plug it in and your agent can actually navigate your machine github.com/wonderwhy-er/D… 2. supervision (roboflow) — ~47.8k★ (+225 today) reusable computer vision toolkit — detection, tracking, annotation, counting. drop-in tools so you don’t rewrite cv boilerplate every project github.com/roboflow/super… 3. speech-to-speech (huggingface) — ~6k★ (+187 today) build fully local voice agents with open-source models. real-time speech-in, speech-out — no api keys needed github.com/huggingface/sp… 4. compound-engineering-plugin (EveryInc) — ~23k★ (+127 today) official compound engineering plugin for claude code, codex, and cursor. structures your ai coding workflow so agents ship better code github.com/EveryInc/compo… 5. agent-governance-toolkit (microsoft) — ~4.8k★ (+78 today) policy enforcement, zero-trust identity, and execution sandboxing for autonomous ai agents. covers all 10 owasp agentic risks github.com/microsoft/agen… one of these saves you a paid tool. bookmark it

@catmanyau totally, the live notes are the real game changer here

@0xProbabillity live meeting notes without that waiting-for-the-transcript lag is the quiet win here

5 open-source AI repos that blew up on GitHub this week → 1. Meetily (Zackriya-Solutions) — ~23k★ (+8.6k this week) fully local AI meeting assistant with live transcription, speaker diarization, and auto-summarization. runs on rust + whisper, zero cloud dependency. replaces otter.ai for free github.com/Zackriya-Solut… 2. OfficeCLI (iOfficeAI) — ~15k★ (+6.5k this week) first office suite purpose-built for AI agents. read, edit, and automate word, excel, and powerpoint from a single binary. no office installation needed. if you’re building agents that touch documents, this is the missing piece github.com/iOfficeAI/Offi… 3. Strix (usestrix) — ~41k★ (+5k this week) open-source AI pentesting tool that finds your app’s vulnerabilities, crafts exploits, and submits PRs with fixes. replaces a $10k+ manual pentest with a docker run github.com/usestrix/strix 4. Herdr (ogulcancelik) — ~16k★ (+4.3k this week) agent multiplexer for your terminal. run claude code, codex, and gemini cli in parallel on the same codebase without them stepping on each other. orchestrates the fleet so you don’t have to github.com/ogulcancelik/h… 5. Voicebox (jamiepine) — ~41k★ (+3.1k this week) open-source AI voice studio. clone any voice, dictate, create — all local. replaces elevenlabs for voice cloning and TTS workflows at $0/month github.com/jamiepine/voic… one of these replaces a paid tool you’re using right now. bookmark it

@TheAIShrink exactly, cost is king. local tools winning on price + privacy. more coming soon

@0xProbabillity 8.6k stars on a local transcription tool this week. developers aren't choosing ideology. cloud got too expensive

exactly! the install story is the secret sauce here. most devs are tired of wrestling with cloud creds, api keys, and brittle dockerfiles just to try something new. these projects win because you can literally docker run or click a single binary and be productive in <60 seconds. that friction drop = explosive growth on github. which one are you planning to try first?

@0xProbabillity zero cloud, single binary, docker run. the install story is doing a lot of work in this list.

@0x0SojalSec this local mcp server is exactly what agents needed zero api costs and full privacy

Your agents can now search, crawl, and research the web locally with zero API costs and nothing leaving your machine. Replaced paid web search APIs with a local MCP server. a local-first MCP server that gives any AI agent proper web search, page fetching, full-site crawling, and structured extraction. Everything runs on your machine, No API keys. No per-request billing, Nothing ever leaves your computer. Multi-engine web search 18 sources reranks results with a local ML model, and aggressively caches everything. Ask the same question twice and it’s instant. The fetch layer even escalates to headless browser when sites try to block it. Works out of the box with Claude Code, Cursor, Codex, Gemini CLI, VS Code, Windsurf, Zed, and Antigravity. Since it speaks plain MCP, any compatible agent can use it. - github.com/KnockOutEZ/wig…

Am I tripping or codex removed 5h limits?

5 open-source AI repos trending on GitHub today → 1. pgrust (malisper) - ~2.2k★ (+774 today) postgres rewritten in rust, now passing 100% of the postgres regression tests. the kind of systems project that makes you double-take github.com/malisper/pgrust 2. stitch-skills (google-labs-code) - ~7.2k★ (+340 today) google's library of agent skills for the stitch MCP server. works with gemini cli, claude code, cursor — one skills repo to rule them all github.com/google-labs-co… 3. claude-code-templates (davila7) - ~29.1k★ (+232 today) cli tool to configure and monitor claude code. templates, presets, and dashboards for power users who live in the terminal github.com/davila7/claude… 4. OpenManus (FoundationAgents) - ~57.2k★ (+226 today) fully open AI agent framework — no walls, no fortress. build autonomous agents with tool use, planning, and multi-agent orchestration github.com/FoundationAgen… 5. rowboat (rowboatlabs) - ~16.2k★ (+112 today) open-source AI coworker with persistent memory. think cursor but for workflows — it remembers context across sessions github.com/rowboatlabs/ro… stitch-skills saves you from wiring agent skills yourself. bookmark it