Model size used to mean RAM size. colibri breaks that link: it runs GLM-5.2, a 744B model, on a 25GB machine with no GPU. Only a small slice is needed per token, so it keeps that in RAM and streams the rest off an SSD. It's slow, a token every few seconds, more a proof the memory wall isn't fixed than a finished tool. 2.1k stars, Apache-2.0. #LocalLLM #LocalInference #MoE #OpenSourceAI

Kimi K3 just released 🔹 2.8 Trillion Parameters, 1 Million Context, Natively  Multimodal🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost🔹 Built for long-horizon agentic coding and self-evolving workflowsKimi K3 is now live on Kimi.com, Kimi Work, Kimi Code, and the Kimi API.Open Weights by July 27, 2026.This a Fable 5, GPT 5.6-sol level modelSome interesting underrated gems in Kimi-K3 release:    > an early K3 wrote the majority of the kernels in the late development stages  > it built a triton-class compiler from scratch, MiniTriton, that delivers performance on par with or better than Triton and torch.compile > then it designed a chip, by a model, for a model, in one 48-hour autonomous run  On many tasks this model performs equivalent to Fable,Opus 4.8  and GPT-5.6 at half or even lesser costExcellent at front end code generation. The first-party API is priced at $3.00/$15.00 per 1M input/output tokens, with cached input discounted 90% to $0.30 per 1M tokens.This is a crazy release from Moonshot.ai

HOLY SHIT. SOMEONE JUST KILLED EVERY SKETCHY VIDEO DOWNLOADER SITE ON THE INTERNET.It's called ReClip. Self-hosted. Open source. Free.Paste a link from YouTube, TikTok, Instagram, X, or 1000+ other sites. Pick MP4 or MP3. Done.No ads. No popups. No trackers. No rate limits.Here's how it works: 👇What ReClip actually does:→ 1000+ supported sites (powered by yt-dlp)→ MP4 video or MP3 audio→ Resolution picker before you download→ Batch downloads. Paste multiple URLs at once.→ Automatic URL deduplication→ Clean web UI in any browser→ Docker supportHere's the part developers will love:Python + Flask backend in ~150 lines.Frontend is vanilla HTML, CSS, JS. No framework. No build step.Two dependencies total: Flask and yt-dlp.That's it. The whole thing is auditable in one sitting.Setup takes one command if you have brew:brew install yt-dlp ffmpeggit clone https://lnkd.in/eYMuXxxjcd reclip && ./reclip.shOpen localhost:8899. You're done.Prefer Docker? One line:docker build -t reclip . && docker run -p 8899:8899 reclip100% Open Source. MIT License. Free forever.If you've ever closed 4 ad popups just to download a 30-second clip, this is for you.

The core mechanism, licensing, keyless design, and benchmark figures are accurate and well-documented across independent sources — this is a genuine, actively discussed project with real technical substance.

The ultimate backbone for AI web agents just hit 69k stars. One framework. Handles everything. Works everywhere. → Single HTTP requests → Concurrent Spider crawls → Cloudflare bypass → CLI extraction → MCP Server What is adaptive scraping? AI agents are smart but their web tools are brittle. They don't know how to handle Turnstile captchas. They crash when a website renames its CSS classes. Scrapling is a framework that teaches them. One library = full web access. Perfectly executed. Every time. This repo handles the hardest parts of the web: → Stealth (TLS fingerprinting, header spoofing) → Parsing (relocates elements when UI changes) → Routing (mix fast HTTP with headless browsers) → Speed (10x faster JSON serialization) Install once: pip install "scrapling[ai]" Then: Connect the MCP server to Claude or Cursor. Your AI agent just got unrestricted, token-efficient access to the internet. GitHub Repo - https://lnkd.in/dCurt4Pc

Every functional claim — the creator, the layout/shortcut mimicry, the license, the install/uninstall simplicity, and the star count trajectory — is accurate and verifiable against primary sources.

 The three-tier routing mechanism, the "Skippy" naming, the Rust/open-source framing, and the consumer-hardware pooling concept are all accurate and specific, matching the project's actual documented design.

A MODEL THAT CLONES A VOICE FROM 20 SECONDS no gpu. no api. no cloud. one pip install ★ 6.8k pocket-tts by kyutai, mit license, pip install pocket-tts - the lab behind moshi: → 100M params, runs on plain cpu - they tried a gpu, it wasn't even faster → ~6x faster than real-time on a macbook air m4, using 2 cores → ~200ms to the first audio chunk, streams the rest → clone a voice from a short sample, or use 26 built-in ones → 6 languages, handles infinitely long text → small enough to run in the browser via webassembly everyone's renting gpu time to make audio. this fits in your pocket https://lnkd.in/dpXEvrZX

 Extremely dense and precise — every mechanism (default-fail gate, fresh-context evaluator, self-maintained handoff, star count, license) is a specific, verifiable claim that matches the actual repository almost word-for-word. This is about as tightly sourced as a one-paragraph post gets.

An AI engineering interview question most people get wrong: How does an LLM choose its next token? The model writes one token at a time. → At each step, it scores every token in its vocabulary. For instance, GPT-4o holds 200,019 tokens. → Softmax turns those scores into probabilities: a distribution to pick from. → The next token is sampled from this distribution. This is 𝘁𝗼𝗸𝗲𝗻 𝘀𝗮𝗺𝗽𝗹𝗶𝗻𝗴. Here are three runtime settings that can change how that pick happens. 𝗧𝗲𝗺𝗽𝗲𝗿𝗮𝘁𝘂𝗿𝗲 Scales the scores before softmax. It reshapes the probability distribution. → T = 1: no scaling. The distribution stays exactly as the model produced it. → Higher T flattens it. Long-shot tokens get a chance, so answers are more varied and creative. → Lower T sharpens it. High-probability tokens dominate, so answers are more focused and predictable. As T approaches 0, sampling approaches argmax. The highest-probability token dominates. ↳ Even temperature 0 may not produce identical outputs every time. ↳ Floating-point operations, batching, and hardware differences can introduce variation. 𝗧𝗼𝗽-𝗽 (𝗻𝘂𝗰𝗹𝗲𝘂𝘀) Adds the highest-probability tokens until their cumulative probability reaches p. The pool grows or shrinks with the model's confidence. → p = 1.0: every token stays in play. → p = 0.9: only the tokens covering 90% of the probability. → p = 0.1: very restricted, close to greedy. 𝗧𝗼𝗽-𝗸 Keeps only the k highest-probability tokens. The pool size is fixed. → k = 50 and the model is confident: you still sample from 50 when only 3 matter. → k = 50 and the model is uncertain: 50 tokens may not be enough. 𝗢𝗻𝗲 𝗿𝘂𝗹𝗲 𝗼𝗳 𝘁𝗵𝘂𝗺𝗯 Temperature and top-p both act on the same distribution. → Change one at a time. → Move both and you stack two diversity controls. The effect gets harder to reason about. Temperature reshapes the curve. Top-p and top-k filter the pool. 𝗧𝗼𝗸𝗲𝗻 𝗦𝗲𝗹𝗲𝗰𝘁𝗶𝗼𝗻 One token is selected from what's left. → After filtering, the remaining probabilities are rescaled to sum to 1. → Higher-probability tokens have a higher chance of being picked. → A weighted random draw picks one token based on those probabilities. The selected token is added to the sequence. Then the process repeats. That is how an LLM chooses its next token. And because the draw can land on a different token each time, the same prompt can produce a different answer. 💾 Save this for your next AI interview ➕ Follow for more AI engineering insights ♻ Repost to help someone in your network

When an AI agent acts in your Slack or GitHub, a gateway holds the login token so the agent never touches it. But the token still leaves on each call, and a redirect to another domain can carry it somewhere it should never go. OpenConnector strips credential headers on any cross-domain redirect and blocks calls bounced to internal addresses. That outbound path is where keys leak. Early repo, 2.1k stars, Apache-2.0.

Your notes do not need another productivity system. They need an interface that can actually reason across them.I have put together a blog that explains how to connect an Obsidian vault with OpenAI Codex and turn a folder of markdown files into an AI-assisted knowledge system. You will learn how to structure your vault, use an `AGENTS.md` file to define rules, set safe permissions, and use Codex to summarize research, connect related ideas, clean up notes, and generate weekly reviews.Why it matters: most second brains are good at storing information but bad at helping you synthesize it. Codex changes that by operating directly across the files you already own, without requiring a complicated integration.This is for researchers, writers, developers, students, founders, and anyone sitting on hundreds of notes they can no longer meaningfully navigate.

6 open-source GitHub repositories for crawling and scraping the web 1./ Firecrawl An open-source web-data API, also available as a hosted service, for searching, scraping, mapping and crawling the web. It can convert individual pages into Markdown, HTML, screenshots or structured JSON, discover URLs across a website, crawl multiple pages and process large URL batches asynchronously. It also supports interacting with pages through actions such as clicking, scrolling and typing before extracting the content. 2./ Crawl4AI A Python crawler and scraper built around producing content for LLMs, RAG pipelines and AI agents. It generates clean or filtered Markdown, supports structured extraction using CSS, XPath or LLMs, and can run deep crawls using strategies such as breadth-first search. Its Playwright-based browser layer can execute JavaScript, preserve sessions, use proxies and handle dynamic or infinite-scroll pages. 3./ Scrapling An adaptive Python scraping framework that supports everything from fetching a single page to running concurrent, multi-session crawls. Its adaptive selector system can save an element’s properties and search for the most similar element when the original selector stops matching after a page redesign. Its spider framework also includes streaming, concurrency controls, pause and resume, and proxy rotation. 4./ Crawlee A JavaScript and TypeScript library for building web crawlers, scrapers and browser-automation workflows. It works with lightweight HTTP and HTML parsing as well as Playwright and Puppeteer for JavaScript-rendered pages. Crawlee provides persistent URL queues, structured-data storage, automatic concurrency scaling, session management and proxy rotation, while leaving the page-specific extraction logic to the developer. 5./ Scrapy A high-level Python framework for crawling websites and extracting structured data. Developers create spiders that issue requests, follow links and parse responses using CSS or XPath selectors. Scrapy’s engine coordinates scheduling and downloading, while item pipelines can validate, transform or store extracted records, and feed exports can write results to formats such as JSON, CSV and XML. 6./ Maxun An open-source, no-code web-data platform that combines extraction, scraping, crawling and web search. Its Recorder Mode captures browser actions and turns them into reusable extraction robots, while its AI Mode uses natural-language instructions for structured extraction. It can also convert webpages into Markdown or HTML, capture screenshots, crawl websites with configurable scope and run locally through self-hosting. ---- "The web is the largest database in the world." [GitHub repo in the comments]

Google just standardized one of the most useful ideas for building AI assistants. It's called the Open Knowledge Format (OKF). → https://lnkd.in/dWdn8ZU2At its core, OKF is a standardized format for building LLM wikis. Instead of storing knowledge behind proprietary APIs, it stores it as Markdown with structured metadata. Because this gives you something that's simultaneously: • Human-readable• LLM-readable• Git-friendly• Vendor-neutral The interesting part isn't the format itself... It's what it enables. Your knowledge becomes a collection of linked documents that both humans and agents can navigate. Each page contains: • Structured metadata (YAML)• Rich documentation (Markdown)• Links to related concepts Instead of loading everything into one enormous context window... Agents progressively navigate the wiki, following links only when they need more information. The same knowledge can then power: • Humans• AI agents• Search• Documentation• Knowledge graph visualizers ...without maintaining separate representations for each. This is a strong signal of where agent memory is heading. Over the past year, I've been leaning toward the same pattern while building my personal AI assistant. Instead of treating memory as something hidden behind a database... Treat it as an LLM wiki that both humans and agents can read, navigate, version, and evolve together. LLM wikis are becoming much more than an implementation detail. They may become the default interface between humans and AI systems. Definitely worth bookmarking if you're building AI assistants, RAG systems, or knowledge platforms. 

Exceptionally dense, specific, and independently verifiable — exact benchmark figures, line counts, hardware configurations, methodology details (cooldowns, gradient-tensor gating), and genuine engineering insights about compiler portability. This is one of the most substantive, first-person technical posts in this batch, backed by a real, citable primary source from the same author.

BREAKING: The open-source AI race just crossed a threshold nobody expected this fast. Moonshot AI dropped Kimi K3, a 2.8 trillion parameter native multimodal model. First open 3T-class release ever. 1 million token context window. 896 experts, 16 activated via a latent MoE architecture. And it lands the same week as GPT-5.6 Sol, Claude Fable 5, and Grok 4.5. The frontier is no longer a closed club. Who loses most if open-weight models keep matching proprietary ones at this scale? PS: AI is creating the new generation of $100M bootstrapped solopreneurs 🔥 This is where you can start: 1. The Claude Code Founder Guide→ https://lnkd.in/ezxNt6yF 2. want to keep your job? read this → https://lnkd.in/e7e3ayuN 3. claude just made outbound fully automatable → https://lnkd.in/ewTepr47 4. The Ultimate AI Fundraising Copilot (Built With Claude) → https://lnkd.in/erTMbAKu 5. The Ultimate Guide to Building Skills for Claude → https://lnkd.in/eVkx4-JS 6. The Prompting Skills Guide → https://lnkd.in/e5Qxxt_a 7. The Garry Tan (YC) Playbook for Building with Claude → https://lnkd.in/ekagtj3d 8. Claude. md Full guide→ https:// lnkd.in/eQfgzppS 9. Understanding Claude Dispatch (Complete Guide) → https://lnkd.in/erfMzRGS 10. Claude For Finance: The CompIete Guide→ https://lnkd.in/egRC5UBa 11. Claude x SEO → https://lnkd.in/eXZSXwKu 12. Anthropic Leaked Product PLAYBOOK: How to 100x Your Coding: → https://lnkd.in/e2mzW_he

Meta just changed the AI pricing war.Not by building the smartest model.By building a frontier-grade AI agent that’s cheap enough to deploy at scale.That’s the bigger story.Meta just introduced Muse Spark 1.1, its first model available through the Meta Model API.This isn’t another chatbot.It’s built for AI agents.Key capabilities:• 1M token context window• Parallel sub-agents• Native tool use• Desktop, browser, and mobile computer use• Long-running autonomous workflowsThe benchmark results are impressive.🏆 Leads on:• MCP Atlas: 88.1• JobBench: 54.7• Humanity’s Last Exam (with tools): 62.1• Finance Agent v2: 57.2Strong coding performance:• Terminal-Bench 2.1: 80.0• SWE-Bench Pro: 61.5• DeepSWE 1.1: 53.3While GPT-5.5, Claude Opus 4.8, and Gemini 3.1 Pro still lead on several coding and multimodal benchmarks, Meta is optimizing for something different:High-performing AI agents at a significantly lower cost.The industry is shifting from:“Which model is the smartest?”to“Which model delivers the best production ROI?”In production, the best model isn’t necessarily the one with the highest benchmark.It’s the one that balances:• Capability• Tool use• Context length• Reliability• Latency• CostThe next AI race won’t be won by chatbots.It will be won by production-grade AI agents that can plan, use tools, collaborate, and execute complex workflows economically.Benchmarks measure intelligence. Production measures value.

𝗚𝗟𝗠-𝟱.𝟮 — 𝘁𝗵𝗲 𝘀𝘁𝗿𝗼𝗻𝗴𝗲𝘀𝘁 𝗼𝗽𝗲𝗻-𝘄𝗲𝗶𝗴𝗵𝘁 𝗰𝗼𝗱𝗶𝗻𝗴 𝗺𝗼𝗱𝗲𝗹 — 𝗻𝗼𝘄 𝗿𝘂𝗻𝘀 𝗼𝗻 𝗮 𝗺𝗮𝗰𝗵𝗶𝗻𝗲 𝘄𝗶𝘁𝗵 𝟮𝟱𝗚𝗕 𝗼𝗳 𝗥𝗔𝗠.744B parameters. 62.1 on SWE-bench Pro — ahead of GPT-5.5. And colibrì runs it on consumer hardware in ~2,400 lines of pure C. No BLAS. No Python. No GPU. Apache 2.0.━━━━━━━━━━━━━The trick:A 744B MoE only activates ~40B params per token — and only ~11GB of those change token to token.→ The dense part sits in RAM — 9.9GB→ The 21,504 experts (~370GB) live on disk and stream in on demand→ A learning cache pins whichever experts your workload actually uses — hit rate climbs from 3% to 98%It gets faster the more you use it.━━━━━━━━━━━━━The honest part, straight from the author:This is not fast. ~0.05–0.1 tok/s on his box. Best community result is ~2 tok/s.It won't replace your coding agent tomorrow. That's not the point.━━━━━━━━━━━━━The point is this:For sparse models, VRAM was never the wall. The per-token working set is tiny — the bottleneck is how fast you feed it.Which turns "can we run a frontier coding model in-house" from a capex question into an I/O question.For any team whose codebase legally can't touch an API — banks, defense, government — that's the whole ballgame.Slow today. The direction is what matters.━━━━━━━━━━━━━

Moonshot AI just released Kimi K3, and it might be the most important open-weight model of 2026. Forget the 2.8T parameters. That number is mostly marketing. Here is what actually matters: 1. Architecture: Kimi K3 is a Mixture-of-Experts (MoE) model with 2.8 trillion total parameters and a 1M-token context window. The two innovations worth understanding are: - Kimi Delta Attention (KDA) Instead of relying entirely on quadratic attention, KDA uses a hybrid linear attention mechanism that dramatically reduces compute and memory overhead on very long contexts. - Attention Residuals (AttnRes) Rather than propagating every hidden representation through all layers, AttnRes selectively reuses useful representations across network depth. 2. Where does it rank? Early independent rankings place it around #4 overall on Artificial Analysis. Current frontier looks roughly like: Claude Fable 5, GPT-5.6 Sol (High), GPT-5.6 Sol, Kimi K3 so it is ahead ahead of Claude Opus 4.8, and GPT-5.5 Moonshot's own evaluations also position K3 ahead of GPT-5.5 and Claude Opus 4.8 on several coding and agent benchmarks, though it still trails Claude Fable 5. As always, wait for more independent evaluations before treating benchmark wins as definitive. 3. Pricing This is where things get interesting (Input, Output) Kimi K3 - $3/M, $15/M GPT-5.6 Sol $5/M $30/M Claude Fable 5 $10/M $50/M For heavy coding agents or enterprise workflows, inference costs can become a major line item. K3 undercuts the flagship proprietary models while staying surprisingly competitive on quality. Try it out yourself on either of these: Kimi Chat, Moonshot API, OpenRouter, Kimi Code, Kimi Work Open-weight downloads are also expected, enabling self-hosting and fine-tuning for organizations that need data sovereignty. If the independent benchmarks hold up, Kimi K3 could become the default starting point for many production AI systems. #KimiK3 #OpenSourceAI #AgenticAI #AIEngineering

A new FFmpeg GUI called Frame just hit 1.5k stars. Instantly, other GUIs feel dated.Here's why it lands differently.Frame is written in Rust on GPUI-CE, a community build of the GPU-rendered UI framework behind the Zed editor. It opens instantly and stays smooth, even during an encode. It feels like a native app because it is one.Marek Jóźwiak (66HEX), a solo dev in Warsaw, got tired of typing FFmpeg flags and built the interface he wished existed.What you actually get:→ Batch a whole folder, then pause, resume, or retry any file→ Stream copy for instant trims and remux, no re-encode→ Full codec and filter control when you want it, H.264 through AV1→ One-click presets for YouTube, TikTok, Reels, X, and Discord→ Burn in subtitles, plus clean or rewrite metadataAnd it's local-first. No uploads, no accounts, no telemetry. Nothing leaves your machine.The honest part: it's young at v0.31.1 and a one-person project, so the odd format will still trip it up.But it's a rare FFmpeg frontend that feels designed, not assembled.693 commits, fully open source. Nice work, Marek Jóźwiak ⭐ GitHub repo => https://lnkd.in/ei_HfpeC--P.S. Need to scale your video transcoding workflow? We have a powerful yet simple platform for that => chunkify.dev