GitHub Models has been retired, everyone: GitHub Models. For those who don't know, it was a service that let you try out free language models for research purposes.
There are many paid models spread across different providers, so testing or evaluating them usually means paying each provider. GitHub Models instead gave access to a limited set of free models.
Recently GitHub has been tightening the costs associated with LLMs, so this shutdown isn't really surprising
The cc-switch project has been updated to allow OpenCode Go to be integrated into Codex—great news!
I previously used maximhq/bifrost as a proxy, but it was unstable and would stop after a short while. After switching back to cc-switch, I’ve had no problems, and the setup is even simpler.
DeepSeek has just released the V4 Flash model. If you’re wondering why it seemed like V4 Flash was launched a while ago, that earlier version was only a preview; the current release is the official one, and the same goes for the current V4 Pro, which is also still a preview.
The latest V4 Flash delivers performance that far outperforms the Pro preview version, and its speed is truly outstanding. V4 Pro will be launched soon, so stay tuned to see how its performance measures up.
I just came across this useful tool, everyone: agentsview.io.
It provides an interface to manage all sessions of most current AI agents such as Claude Code, Codex, OpenCode, Pi, etc. You can view the list, filter, and see very detailed statistics for each conversation or for all of them. This supplies useful information about how often and to what extent AI agents are used, serving as a basis for many future activities.
I haven't posted lately because I've been focusing on working on OpenNotas, everyone. Stay tuned for the upcoming updates of this personal note‑taking app.
Google has lately seemed to "prefer" releasing Lite versions rather than Pro ones. Yesterday it announced Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. The Lite models give us an affordable price and fast speed, handling everything from simple to complex tasks.
What people are really hoping for, though, is a Pro version that can truly compete with Fable 5 😌
It looks like Claude Code has already started rolling out the Rust‑rewritten version of Bun to production for millions of users 😱
Someone on the internet created codex-resets.com to track, record, and announce whenever Codex resets its 5‑hour‑per‑week usage limit. As you know, a “reset” means restoring the AI model’s usage limits so users can keep using it.
It looks like OpenAI’s Codex has been resetting those limits a lot recently—every time there’s good news it gets "reset" which is why this site exists. Hopefully Claude Code will do the same
Kimi K3 has officially launched. It’s a 2.8‑trillion‑parameter model built on Kimi Delta Attention (KDA) and Attention Residuals, able to understand images and offering a context window of up to 1 million tokens. It’s the world’s first open‑source model in the 3‑trillion‑parameter class, designed for advanced AI scenarios such as programming, knowledge bases, and reasoning.
The price is pretty high, $3 / $15 per Input / Output (comparable to Sonnet 5), although its performance is being compared to GPT 5.6 or Fable 5. Kimi modestly admits the performance isn’t yet on par with competitors, but overall it provides good value for the money, and it’s open source.
Luckily, my Go plan doubled the quota for using Kimi K3 this weekend, so I gave it a try. After only 4–5 prompts I had almost exhausted the five‑hour limit 🥶
Không ổn rồi mọi người ạ, cho nó sửa code phát là ăn limits luôn, limits tuần + tháng cũng tăng chóng mặt 😵💫
Trải nghiệm thì chưa prompt được nhiều nên chưa đánh giá được. Nhưng mà tốc độ suy luận khá chậm, suy luận nhiều, code viết ra format chưa đồng nhất, đặc biệt là đôi lúc còn sai chính tả tiếng Việt 🥶
This morning I came across this really interesting news, everyone. Telegram Serverless lets you run backend code for bots and Mini Apps directly on Telegram’s infrastructure. You just write simple JavaScript modules, deploy them with a single command, and Telegram runs them in a sandbox environment.
Normally you need a server to implement bot logic, but with a server‑less approach you no longer have to worry about infrastructure. As you know, Telegram bots are very useful and have many practical applications—like the several channels I run that serve as handy information‑sharing and monitoring hubs.
The only downside is that it’s still in closed beta, so it’s not usable yet. We’ll just have to wait 😅