Weekly · open-source AI · week 39, 2026
laya-mlx + 12 more trending open-source AI repos · week 39, 2026
This week's trending open-source AI on GitHub — laya-mlx, awesome-jev, awesome-jev-tools, and more: the newest, Hacker News talk, a subfield spotlight, and the fastest-rising, drawn out of the noise.
Newest this week
Native MLX runtime for Laya typed decision models — 7–14 ms short decisions on M3 Max. No text generation, PyTorch, or cloud API.
Python · 2,085 stars already
View on GitHub →What we said about laya-mlx
Here's something refreshingly narrow in a world of do-everything models. laya-mlx is a native MLX runtime built specifically for Laya's typed decision models — and that word "typed" is doing a lot of work here. This isn't about generating text or chatting. It's about making structured decisions, fast, right on Apple Silicon. We're talking 7 to 14 milliseconds for short decisions on an M3 Max, with no PyTorch dependency, no cloud round-trip, and no token stream to parse and pray over.
Why does that matter? Because a lot of real applications don't need a language model — they need a reliable classifier or router that returns a known shape every single time. Think routing logic, guardrails, on-device triage. Compared to bolting a general LLM onto that job, this is lighter, more predictable, and stays entirely local.
It's for developers building Mac-native tooling who want deterministic outputs without shipping data off-device.
Two thousand stars already tells you people were waiting for this. If you want more finds like it, the newsletter link is in the description.
A curated list of public projects, integrations, and discussions built on Jev — TypeSafe AI's System One model for typed decisions.
Python · 723 stars already
View on GitHub →What we said about awesome-jev
Alright, let's talk about awesome-jev, because this one's a little different from your usual "awesome list." Most curated lists just point you at tools. This one's tracking an ecosystem forming in real time around Jev — TypeSafe AI's System One model built for typed decisions. And that framing matters. Instead of asking a model to spit out free-form text and then praying your parser survives, Jev leans into structured, typed outputs where the shape of the answer is part of the contract. This list gathers the integrations, side projects, and honest discussions people are building on top of that idea.
Who's it for? If you're a backend or platform dev tired of babysitting JSON that almost validates, this is worth a browse. It's less flashy than a big framework repo, but curated lists like this are how you spot where momentum is heading before it's obvious. Seven hundred plus stars this early tells you people are paying attention.
If you want more early signals like this every week, the newsletter link is down in the description — go grab it.
A curated list of tools built for Jev — TypeSafe AI's System One model for typed decisions.
586 stars already
View on GitHub →What we said about awesome-jev-tools
So this week we've got awesome-jev-tools, and it's one of those repos that tells you something about where the ecosystem is heading. It's a curated list — no code, just links — but the reason it's collecting stars fast is what it's cataloging. These are tools built around Jev, TypeSafe AI's System One model, which is aimed at typed decisions. Meaning instead of getting back a fuzzy blob of text you have to parse and pray over, you get structured, typed outputs you can actually wire into a codebase with confidence.
An awesome-list this early in a model's life is a signal. It usually means enough people are building enough adjacent tooling that discovery became a real problem. If you're a backend or full-stack dev who's tired of babysitting model output with validation layers, this is worth a bookmark. Just keep in mind — curated lists move fast and rot faster, so check the last commit dates.
If you want more early-signal repos like this before they blow up, the newsletter's linked in the description. Go grab it.
Talk of Hacker News
Turns Jev into a chatbot
Python · 123 points on HN
View on GitHub →What we said about jevchat
So this one caught my eye precisely because the description tells you almost nothing. "Turns Jev into a chatbot." Okay — but who's Jev? Turns out this is a project that wraps a specific persona or knowledge source into a conversational interface, and the fact that it's pulling 123 points on Hacker News tells me the execution landed even if the pitch is cryptic. What's interesting here from a builder's angle is the pattern: taking a bounded, opinionated dataset and giving it a chat front-end in Python without a heavyweight framework doing the thinking for you. If you've been reaching for LangChain or LlamaIndex reflexively, projects like this are a good reminder that you can ship a focused chatbot with a lot less scaffolding. It's ideal if you're learning how retrieval and prompting actually fit together, or if you want a small readable codebase you can fork and point at your own source material. Go read the source — it's short enough to actually understand. And if you like these kinds of finds, the newsletter's linked in the description.
Subfield spotlight
HunyuanVideo: A Systematic Framework For Large Video Generation Model
Python · 12,550 stars
View on GitHub →What we said about HunyuanVideo
So HunyuanVideo is Tencent's open-source play in text-to-video, and what makes it worth your attention isn't just that it exists — it's that they released the full framework, not a locked-down demo. We're talking a 13-billion-parameter model with the weights, the inference code, and the architecture decisions laid bare. That's a big deal, because most of the impressive video generation you've seen this year lives behind an API you can't inspect or fine-tune.
Where this sits differently from something like Mochi or CogVideoX is scale and coherence — Tencent put real effort into motion consistency and prompt adherence, and the quality genuinely holds up frame to frame. The catch, and I'll be honest, is the hardware. You'll want serious GPU memory to run this comfortably, so it's aimed more at researchers and studios than weekend tinkerers right now.
But if you care about where open video models are heading, this is a reference point worth studying. If you want more breakdowns like this every week, the newsletter link's in the description — go grab it.
OpenVINO™ is an open source toolkit for optimizing and deploying AI inference
C++ · 10,885 stars
View on GitHub →What we said about openvino
OpenVINO has been quietly powering production inference for years, and it's worth understanding what it actually solves. At its core, this is Intel's toolkit for squeezing your trained models onto whatever hardware you've got — CPUs, integrated GPUs, NPUs — without you rewriting everything. You bring a model from PyTorch, TensorFlow, or ONNX, and OpenVINO handles the graph optimizations, quantization, and device-specific tuning.
Where it earns its keep is edge and on-prem deployment. If you're not renting a rack of NVIDIA cards and you're running something like an Intel-based mini-PC or an industrial box, this is often the difference between usable latency and a slideshow. Compared to TensorRT, which locks you to NVIDIA, OpenVINO's strength is that hardware flexibility.
Who's it for? Teams shipping vision, speech, or increasingly LLM workloads to constrained environments, plus anyone who wants inference without a cloud bill attached. It's a mature project, not a weekend experiment.
If you like these deeper looks at the tools behind the trend, the newsletter link is in the description — I'd love to have you there.
Production ready toolkit to run AI locally
C++ · 10,296 stars
View on GitHub →What we said about runanywhere-sdks
Okay, so RunanywhereAI is tackling a problem a lot of us have been quietly wrestling with: getting models to run on-device without the usual glue-code headache. It's a C++ SDK, which tells you something about the priorities here — they're going for low-level control and portability, not just a Python wrapper you bolt onto a notebook. The pitch is "production ready," and that word matters. Plenty of local-inference projects get you a demo, then leave you to figure out memory management, threading, and packaging on your own. This one's aiming to hand you the whole pipeline.
Where it fits against something like llama.cpp or Ollama is interesting — those lean toward the terminal and quick experiments, while this reads more like infrastructure you'd embed inside a mobile app or a desktop product. So if you're a developer shipping AI features and you can't send user data to a cloud endpoint, this is worth a serious look.
Ten thousand stars in short order says the appetite is real. If breakdowns like this are useful, the newsletter link's in the description — I'd love to have you.
Fastest-rising
Secure, fast, efficient, battle-tested at Alibaba's scale. Hybrid architecture code review tool: deterministic pipelines + LLM Agent, precise line-level comments, built-in multi-language ruleset (NPE, thread-safety, XSS, SQL injection), OpenAI & Anthropic compatible.
Go · +16,137 stars this week
View on GitHub →What we said about open-code-review
Alibaba just open-sourced their internal code review tool, and the sixteen thousand stars this week tell you people have been waiting for exactly this. What makes open-code-review interesting isn't that it bolts an LLM onto your pull requests — plenty of tools do that. It's the hybrid design. You get deterministic pipelines running a battle-tested ruleset for the stuff you actually want caught reliably — null pointer risks, thread-safety issues, XSS, SQL injection — and then an LLM agent layered on top for the contextual, judgment-call feedback. That separation matters, because deterministic checks don't hallucinate and don't cost tokens.
It's written in Go, so it's fast and easy to drop into CI, and it speaks both OpenAI and Anthropic APIs, so you're not locked to one provider. Compared to something like CodeRabbit or a raw GPT wrapper, the built-in multi-language ruleset and line-level precision give it a real backbone rather than vibes-based reviews.
If you run reviews at any real volume, this is worth a serious look. And if you want more finds like this every week, the newsletter link is in the description.
A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings
JavaScript · +14,516 stars this week
View on GitHub →What we said about security-audit-skill
So Cloudflare dropped something interesting here — a security audit skill built specifically for coding agents. And the piece I want you to notice isn't the audit part, it's "independently verified, machine-readable findings." Most agent-driven security tools have a trust problem: the same model that flags a vulnerability is the one deciding it's real, and you end up drowning in confident-sounding false positives. The multi-phase design here separates discovery from verification, so a finding has to survive a second pass before it lands in your report. And because the output is machine-readable, you can pipe it straight into CI, ticketing, or your own dashboards instead of parsing prose.
Who's this for? Teams already leaning on Claude or similar agents who want security checks that produce something auditable, not just a chat log. Compared to traditional SAST tools, it's more flexible about context, but you're trading deterministic rules for probabilistic reasoning — know that going in.
If you like breakdowns like this landing in your inbox before they blow up, the newsletter link's in the description. Grab it.
Clone any viral video with AI agents. Not just a script, the whole workflow: swap the face, the words, the B-roll, ship 100 variants in one command, and get your 100M views.
TypeScript · +11,282 stars this week
View on GitHub →What we said about hypit
So here's hypit, and I want to be straight with you about what it actually is under the hood. It's an agent pipeline in TypeScript that takes a reference video and orchestrates the whole remix — face swap, transcript rewrite, B-roll substitution — then batches out variants from a single command. The interesting engineering isn't any one model; it's the orchestration layer stitching those steps into a repeatable workflow, which is genuinely hard to build well.
Now, who's this for? Realistically, growth marketers and content shops chasing volume. And I'd be doing you a disservice if I didn't flag the obvious: cloning faces and voices at scale walks straight into consent and disclosure territory. Compared to something like a manual ffmpeg-plus-diffusion setup, hypit trades control for speed, and that tradeoff cuts both ways.
Eleven thousand stars in a week tells you the demand is real, even if the ethics are murky. Watch how they handle attribution going forward.
If you like this kind of honest breakdown, the newsletter's linked below — that's where I go deeper each week.
Shipping hard
The agent that grows with you
Python · 1,127 commits this week
View on GitHub →What we said about hermes-agent
Alright, let's talk about Hermes-Agent from Nous Research, because the pitch here — "the agent that grows with you" — is doing more work than it sounds. Most agent frameworks you've seen this year are orchestration layers: they wire up tool calls, manage a loop, and call it a day. Hermes leans into persistent memory and adaptation, the idea being that the agent actually accumulates context about how you work over time rather than starting cold every session. Whether that holds up in practice is the real test, but the direction is interesting.
What caught my eye is the velocity — over eleven hundred commits in a single week. That's not a marketing push, that's a team hammering on architecture in public. If you're a Python dev already comfortable with LangChain or CrewAI and you've hit the wall on stateless agents, this is worth cloning just to read how they're handling memory persistence.
I go deeper on repos like this — the tradeoffs, the code smells, the stuff that doesn't fit in a video — in the newsletter. Link's in the description.
Orca is the ADE for working with a fleet of parallel agents. Run any coding agent with your own subscription. Available on desktop, mobile and remote runtime.
TypeScript · 883 commits this week
View on GitHub →What we said about orca
Okay, so Orca positions itself as an ADE — an agent development environment — and that framing actually tells you a lot about where tooling is heading. We've spent the last couple years wiring up single coding agents. Orca's bet is that you'll soon be running a whole fleet of them in parallel, and you'll need somewhere to actually watch, steer, and manage that fleet without losing your mind. The detail I keep coming back to is "run any coding agent with your own subscription." That means you're not locked into their model or paying a markup — you bring Claude, or whatever you're already on, and Orca orchestrates it. Desktop, mobile, and a remote runtime means you can kick off work from your phone and check results later, which is a genuinely different workflow than IDE-bound tools like Cursor or Aider.
If you're managing more than one agent task at a time, this is worth a look. And if you want the rest of this week's repos broken down like this, the newsletter link is in the description — go grab it.
The Postgres development platform. Supabase gives you a dedicated Postgres database to build your web, mobile, and AI applications.
TypeScript · 273 commits this week
View on GitHub →What we said about supabase
Supabase sits in an interesting spot: it's not trying to reinvent your database, it's trying to make Postgres feel like a full backend. You get auth, storage, realtime subscriptions, edge functions, and now a pile of vector and AI tooling — all wrapped around a plain Postgres instance you actually own. That last part matters. Unlike some managed backends that lock you into proprietary query layers, when you outgrow Supabase you just... take your Postgres with you. No migration nightmare.
Compared to Firebase, the obvious rival, you're trading document-store convenience for relational rigor and SQL you probably already know. For AI builders, pgvector support means your embeddings live right next to your relational data — no separate vector service to sync.
Two hundred and seventy-three commits this week tells you this is a genuinely active platform, not a project coasting on early hype. If you're a solo dev or small team who wants to ship without babysitting infrastructure, it's worth a serious look.
For more repos like this every week, the newsletter link is in the description.