Practical AI Guides & a Free Claude Code Course

Recursive describes Flint in its own words as "Recursive's mixture-of-experts language model for coding agents, with 39.8 billion total parameters and approximately 2.5 billion active per token." It is built to dig through a codebase, fix bugs, add features, and write tests, and it can hold a long back-and-forth while calling outside tools, which is what an agent needs to actually do work rather than just talk. In Recursive's own evaluations, it scores 74.2% on SWE-bench Verified and 89.7% on SWE-bench Pro v2, both benchmarks built from real bugs in open-source projects. Unsloth, a team known for shrinking large models down to run on ordinary hardware, packaged it into ready local files; of the compressed options, the one unsloth labels Q4_K_XL runs about 25 GB, and a much smaller but noticeably weaker one runs about 11 GB. Because only a fraction of its parameters activate per token, it does not need to load the full 40 billion into memory at once, which is why it fits on a single well-equipped machine at all.

Black Forest Labs is a German team behind the whole FLUX family, one of the best-known open image-generation models. The authors describe FLUX 3 Image Klein as "an open model with 5.6 billion parameters for text-to-image generation and image editing." You write a text description and get a picture, or go further and hand it up to ten of your own photos, and it builds one new scene pulling details from each. This is the lightweight, ready-to-run version of the FLUX 3 release — the other version, Base, is meant for fine-tuning and research, while Klein is for people who just want to install it themselves. The license here is different from normal open-source code, so read the terms yourself before selling anything made with it.

VictorTaelin/uniichat

UniiChat is an AI agent that has a single chat with no “new chat” button. You can chat with it for days, weeks, or even years! Each message gets compressed down to a single line. These lines get further compressed into shorter lines over time. Information gets lost, but never deleted. When the agent wants to recall past information, it can expand a short line back into its initial message. This compression happens in the background using a second, cheaper AI model as you type your response. The same chat can be accessed both via a web browser and via a command line interface (CLI) tool named unii. In addition, on macOS, the agent is able to interact with the screen in a special mode. The agent is able to spawn background sub-agents that perform tasks and then send their results back to the main chat. The agent is able to import historical chats from Claude Code (Anthropic’s coding tool), Codex (OpenAI’s coding tool), or export conversations from ChatGPT. To install UniiChat, simply ask your agent to deploy UniiChat from github.com/VictorTaelin/uniichat following the instructions in its README (i.e., the project’s “setup guide”). Note that you will need Bun (a high performance JavaScript runtime) and your own Anthropic or OpenAI API key in order to run the server. This runs on a regular laptop without a GPU.

CopilotKit/OpenDots

OpenDots is a free template from CopilotKit, a company that builds AI assistant tools. It's for always-on AI 'coworkers' that move between chat, phone calls, and Slack, a workplace messaging app. Your documents live in 'Spaces' — shared workspaces. Each agent is called a 'Dot.' Every Dot has its own name, role, instructions, and list of allowed tools. Any Dot can get its own isolated computer — a separate browser and file system. This uses OpenBot, another CopilotKit project. Its browser profile and files are kept between restarts. Before a Dot saves a page or touches a real account, it shows a card with 'approve' and 'decline' buttons. It waits for your decision. This covers things like email, a calendar through a service called Composio, or any tool with write access. All chats and agent work are stored in CopilotKit Intelligence. It comes as a paid cloud service, a local test version, or a separately licensed self-hosted option. The local version runs in Docker, a tool for running software in isolated containers. The authors call it a starting point for building, not a finished product. It's in alpha, an early test stage. Some features, like calls and handing work off to Slack, are only tested locally so far. To install it, ask your coding AI agent to deploy OpenDots from github.com/CopilotKit/OpenDots. You'll need Node.js 24, a JavaScript runtime, and npm. You'll also need a CopilotKit Intelligence account and a model key, such as OpenAI's. Then have it run npm run dev and open http://127.0.0.1:5173. It runs on a regular laptop — no graphics card needed.

cheaper-inference/OmniRoute

OmniRoute is a free, open-source gateway (a program that routes your requests) you install on your own computer, giving you one local address instead of setting up each AI tool separately. Behind that address are 375 AI model providers, more than 150 of them with free access you sign up for yourself under their own terms, and the gateway can automatically switch between them if one is down or capped. It also compresses: a chain of filters shrinks, on the fly, the text that normally eats tokens for nothing — repeated command output, logs, long tool replies — and the project's own documentation claims this saves 15 to 95% of tokens per request, while leaving code, links, and structured data untouched. You connect it to regular coding agents and editors — Claude Code (an AI coding tool), Codex, Cursor, and others — through that same local address; it's also available as an MCP server (a way for an AI agent to control other tools), so the agent itself can manage the gateway: add providers, change compression settings, check status. Per the project, requests go straight to the chosen provider with no OmniRoute cloud server in between, and stored keys are encrypted. To install: ask your agent to set up OmniRoute from github.com/diegosouzapw/OmniRoute, following the project's own instructions page.

Meta has released their new open-source (downloadable and runnable) model, Muse Spark Open. It’s an 816 billion parameter model with 42 billion effective params per token (word piece) generated (MoE). This means less computation everywhere. With a context size of 1 million tokens, this model can process a lot of information at once (conversation/code). According to the paper, this model was designed for code/agent tasks but can also understand text/image/video/audio and respond back in text. They report a score of 76.4 on Terminal-Bench 2.1 (an agent benchmark in a terminal environment, i.e. command line interface). This benchmark hasn’t been verified by Hugging Face (model hosting service). Licensing is under the Muse Model License which allows you to deploy this model in your application without paying Meta, unless you or your affiliates generate more than $20M in revenue annually, then you will have to negotiate a separate contract with Meta. Play around with this model here in HuggingChat (no hardware required) by visiting https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B, giving it an agent task in a terminal and watching it reason through things and write some code.