This week's GitHub landscape is an absolute firehose of interesting projects — from DeepSeek making moves on LLM inference optimization, to AI agents placing sports bets on-chain, to a handful of genuinely useful dev tools I'm already running. Let's dive in.

DeepSeek Drops DeepSpec: Speculative Decoding Goes Full-Stack

The big headline this week is DeepSeek-ai/DeepSpec — a full-stack codebase for training and evaluating speculative decoding algorithms, and it rocketed to 3,500+ stars in just three days. If you've been following LLM inference, you know speculative decoding is one of the most promising techniques for speeding up generation without sacrificing output quality. The idea is simple: use a small draft model to predict tokens, then have the big model verify them in parallel. But implementing it from scratch is a beast.

DeepSpec changes that. It gives you the entire pipeline — training, evaluation, and deployment — all in one place. Built in Python, it's designed for researchers who want to iterate on decoding strategies without rebuilding infrastructure every time. For anyone working on production LLM serving, this is one to watch.

BenchFlow Launches awesome-evals: The Anti-BS Agent Evaluation List

One of the freshest repos this week — launched June 24th — is benchflow-ai/awesome-evals. It's already at 584 stars and climbing fast. The tagline says it all: "A curated, non-BS library of the best resources for building and evaluating AI agents." Papers, blogs, talks, tools, benchmarks — all in one place.

Why does this matter? Because agent evaluation is currently the Wild West. Everyone's building agents, but figuring out whether they actually work is still fragmented across a dozen frameworks with zero interoperability. awesome-evals is trying to be the Rosetta Stone. If you're building agents with any seriousness, bookmark this one.

Eleven AI Agents Are Out Here Gambling (On-Chain)

Okay, this one is just wild. winsznx/theeleven — 714 stars in four days — deploys eleven autonomous AI agents that open live football prop betting markets on X Layer. We're talking a custom Uniswap v4 hook, gasless USDT0 staking, and the whole thing runs on an MCP server. It's absurd, technically impressive, and exactly the kind of thing that makes you stop and think "oh, we're living in the future now." The TypeScript + Solidity stack is well-organized, and the architecture docs are refreshingly clear for a project this complex. Even if you don't care about sports betting or DeFi, the agent coordination patterns here are worth studying.

pdfx: What If PDFs Could Hold Multiple Files?

AlexandrosGounis/pdfx crossed 545 stars on launch week, and it's one of those ideas that makes you smack your forehead. It extends the PDF standard to let you store multiple files inside a single PDF via metadata. Think of it like a ZIP file, but natively in the PDF spec. No extraction needed — the files live in the document's metadata layer.

Practical use cases? Imagine bundling a report PDF with its source data, or embedding images alongside a design document. It's TypeScript-based, the API is clean, and it already handles the edge cases around reader compatibility gracefully. I love projects like this — small scope, huge impact.

mpatch: Context-Aware Diffs for the AI Era

If you've ever tried to apply AI-generated code patches and watched them fail because the line numbers shifted by two, Romelium/mpatch is for you. Written in Rust (naturally), mpatch applies diffs based on context rather than line numbers. It does fuzzy matching, handles Markdown-formatted diffs, and is designed specifically for the LLM code-generation workflow.

Only 18 stars right now, but this is a sleeper hit. Anyone working with AI-assisted development workflows knows the pain of patch application failures. mpatch solves that problem at the system level rather than forcing you to work around it in every tool.

Poirot: A macOS Companion for Claude Code

a7t-ai/poirot (190 stars) is a native macOS app that gives you a session viewer and transcript explorer for Claude Code. You can browse past sessions, explore diffs, and re-run commands without digging through terminal history. It's built with SwiftUI, uses Sparkle for updates, and is available via Homebrew cask.

This scratches a real itch. Claude Code sessions generate a lot of output, and scrolling back through terminal history to find that one prompt you ran three hours ago is not productive. Poirot makes it browsable, searchable, and actionable. Pair this with mpatch and you've got a genuinely solid AI dev workflow.

flemma.nvim: AI Conversations Where Every Document is Yours

Flemma-Dev/flemma.nvim describes itself as "an AI workspace inside Neovim where every conversation is a document you own." That's a philosophy I can get behind. It supports Claude, ChatGPT, Gemini, and Vertex AI — all inside Neovim buffers. Conversations are just files. You own them, you version them, you grep them.

Most AI chat tools trap your conversations inside their database. flemma.nvim puts them in your filesystem where they belong. 55 stars and growing. For the Neovim crowd, this is a natural fit.

torlink: Zero-Setup Torrents in Your Terminal

And finally, something purely practical. baairon/torlink (436 stars, launched June 25th) is a sleek, zero-configuration torrent finder and downloader that lives in your terminal. TypeScript-based, works with magnet links, requires no setup. You install it, you run it, it works. No daemon. No config file. No web UI. Just the terminal.

Sometimes the best tools are the ones that get out of your way. torlink does exactly that.

Wrapping Up

This was a strong week for open source. DeepSpec gives researchers a proper playground for speculative decoding. awesome-evals is bringing order to the agent evaluation chaos. The Eleven is pushing what's possible with multi-agent systems and DeFi. And tools like mpatch and Poirot are quietly making the AI developer experience better one patch at a time.

Clone something. Break something. Build something.

— Allan Ali