Six AI tools launched this week. Four are open-source and free to run today. This covers what each one does, how a business owner actually uses it, and the limits to know before committing any workflow to it.
Most AI news weeks are announcements followed by a waitlist. This week was different.
Six tools shipped. Four are on GitHub right now. One runs a 2.8 trillion parameter model on a graphics card with under 4 gigabytes of RAM. One gives automation bots a private browser with their own files and logins. One sandboxes your agents in an isolated desktop so they can work without touching your live systems.
If you track AI tooling costs and capabilities for your business, this is a week worth reviewing.
Why These Releases Matter for Business Owners
The recurring argument for paying for proprietary AI access has been capability. Open-source models were not competitive. That argument is narrowing fast.
GLM 5.3, released this week by Zhipu AI, is now the highest-scoring open model on Terminal Bench, a benchmark that tests coding and systems tasks. It beats Anthropic and OpenAI on cybersecurity. It is free.
Separately, airLLM runs a 2.8 trillion parameter model on a standard graphics card with under 4 gigabytes of VRAM. It does this by streaming one model layer at a time. No expensive server required. Free and open-source.
Neither of these releases means you should cancel your API subscriptions today. They mean the business case for proprietary AI needs to be re-evaluated more frequently than it did a year ago. The gap is closing. This week's releases are the evidence.
Which Tool Fits Which Business Problem
| Tool | Best for | Setup time | Cost |
|---|---|---|---|
| GLM 5.3 | Coding, security review | Hosting-dependent | Free |
| airLLM | Local LLM inference | 1-2 hours | Free |
| OpenBot | Browser automation | 2-4 hours | Free |
| OpenSandbox | Agent task isolation | 1-2 hours | Free |
| code-review-graph | Cutting AI token costs | 15 minutes | Free |
| Replit Free Mode | Dev work on Replit | None (if subscribed) | No extra cost |
Use the table to pick your starting point. Then read the relevant section below before you install anything.
GLM 5.3: The Open-Source Model That Topped the Benchmark
What it is. GLM 5.3 is a large language model from Zhipu AI. It is open-source. As of this week, it is the highest-scoring model on Terminal Bench, a benchmark focused on coding and terminal-based tasks. It beats both Anthropic's and OpenAI's current models on cybersecurity benchmarks. Zhipu reports a 50% improvement in coding performance compared to the prior version.
How a business owner uses it. Development work, code review, and security-adjacent tasks are the strongest starting points. If your team uses AI for any code-touching operations, GLM 5.3 is the open-source alternative that is now competitive with paid APIs.
Self-hosting removes per-token billing entirely. It also means no data leaves your infrastructure. For businesses with sensitive code or regulated data environments, that tradeoff is worth serious evaluation.
Where to find it. Visit zhipuai.cn/en. Confirm your server infrastructure can handle the model size before committing anything.
What you should know before switching. Benchmark performance on Terminal Bench does not predict performance on your specific tasks. Run GLM 5.3 alongside your current model on actual work before changing any live workflow. Benchmarks favor the task types they test. Your operations may favor different ones.
airLLM: Run a 2.8 Trillion Parameter Model on Consumer Hardware
What it is. airLLM is an open-source library that makes it possible to run very large language models on consumer-grade GPUs. It works by loading one model layer at a time rather than requiring the whole model to fit into memory at once.
What makes it notable. A 2.8 trillion parameter model on under 4 gigabytes of VRAM. The layer-streaming approach removes the hardware requirement that has historically locked models at this scale to expensive server clusters. The software is free.
How a business owner uses it. If your team has hit a ceiling on cloud inference costs, airLLM is the path to running frontier-scale models on hardware you already own. No per-call cost. No API dependency. The model runs locally, which also means no data transmitted to external servers and lower latency for on-premises workflows.
Batch processing, document analysis, and non-time-sensitive AI tasks are the best starting points for local inference at this scale.
Where to find it. The repository is at github.com/lyogavin/airllm. Installation uses standard Python tooling.
What you should know before switching. Layer-by-layer loading is slower than standard inference. You trade speed for hardware efficiency. Test actual latency against your workflow requirements before committing. If your use case requires fast turnaround, this tradeoff may not work for you.
OpenBot: A Self-Hosted Automation Bot That Runs Its Own Browser
What it is. OpenBot is an open-source alternative to hosted automation bots, built by CopilotKit. It runs on your own infrastructure. Each bot instance gets its own browser, its own file storage, and its own login credentials. You watch it work in real time. You take over control whenever you want.
How a business owner uses it. Any browser-based work done manually today is a candidate: prospect research, form submission, data collection from web sources, report generation from online dashboards. OpenBot handles these tasks using a browser, the same way a person does, without needing the source site to provide an API.
The observable-and-interruptible design matters in practice. This is not a black-box automation. You can watch the session, confirm the bot is doing the right thing, and take the keyboard if it is not. That level of control is what makes it usable for high-stakes tasks.
Multiple bots can run in parallel with isolated credentials, so one bot logging in to one account does not conflict with another.
Where to find it. The repository is at github.com/CopilotKit/openbot.
What you should know before switching. You own the maintenance. Updates, security patches, and infrastructure monitoring are yours to manage. Budget 2-4 hours for initial setup and ongoing attention for anything running in production. This is not a set-and-forget tool.
OpenSandbox: The Isolated Workspace Your Agents Were Missing
What it is. OpenSandbox is an isolated execution environment for AI agents, released by Alibaba as open-source. Inside the sandbox, an agent can run code, browse the web, control a full desktop, and use developer tooling including Cursor, Codex, and Gemini CLI.
How a business owner uses it. Agents that can browse, execute, and interact with software do more useful work than agents that only output text. The missing piece has been containment: a place where an agent can take real actions without touching your live systems or production data.
OpenSandbox fills that gap. Run financial models, test automations, prototype workflows. The agent works inside the sandbox. Your actual systems stay outside it. You review the output, then decide whether to run it in production.
Where to find it. The repository is at github.com/alibaba/OpenSandbox.
What you should know before using it in production. This project is early-stage. Before using it for anything connected to real accounts or live data, verify what can and cannot exit the sandbox environment. Early-stage containment tools sometimes have undocumented gaps. Read the security documentation first. Do not build production dependencies on it until you have confirmed the boundary behavior on your specific setup.
code-review-graph: Cut AI Token Costs on Your Codebase by Up to 4x
What it is. code-review-graph maps your codebase into a set of linked markdown files. Your AI coding tool reads the map instead of raw source files. The reported result: up to 4 times fewer tokens consumed per interaction with your codebase. Two commands to install. Free.
How a business owner uses it. Token costs and context limits are real constraints when running AI coding assistance on any non-trivial project. code-review-graph addresses both without changing your workflow. You generate the map once. Your existing AI tools read from it automatically.
Compatible with Claude, Codex, and other AI coding environments that accept file context.
Where to find it. The repository is at github.com/tirth8205/code-review-graph.
What you should know before relying on the number. The 4x reduction is a reported figure, not guaranteed across all codebases. Your actual improvement depends on codebase structure and how your AI tool processes file context. Test on your specific repository. Larger, more interconnected codebases tend to benefit most. Small projects with few files may see smaller gains.
Replit Free Mode: Lower AI Costs on a Platform You Already Use
What it is. Replit added a free AI tier to paid plans, powered by GPT-5.6 Luna. Lightweight AI tasks within Replit now run without consuming your credit balance.
How a business owner uses it. If your team already uses Replit for development or prototyping, lightweight AI tasks now cost less. This is a cost reduction inside an existing tool, not a new capability. Its value depends entirely on how much lightweight AI usage your team runs through Replit.
What you should know before adjusting your usage. This applies to existing Replit paid plan subscribers. Lightweight-task classification determines what is free. Heavier AI usage still draws credits. Confirm the classification criteria in Replit's current documentation before changing how your team uses the platform.
Honest Limits You Should Know Before Switching Anything
Open-source tools require you to run them. A hosted API costs money per call. A self-hosted model costs server capacity, setup time, and ongoing maintenance. Neither is cheaper in every dimension. The right choice depends on your team's capacity to manage infrastructure.
Benchmark scores are not use-case scores. GLM 5.3 topping Terminal Bench tells you it performs well on that benchmark's task set. It says nothing about how it performs on your invoices, your client emails, or your edge cases. Test before switching.
Several of these projects are early-stage. Early-stage open-source projects ship with incomplete documentation, changing APIs, and bugs that have not been found yet. Factor that into any decision to build production workflows on top of them.
The most important limit is not about any specific tool. No amount of free infrastructure replaces knowing which problems in your business are worth automating. A well-run process benefits far more from automation than a broken one. Fix the process first. Then automate it.
One Thing to Do Before This Week Ends
Pick one tool from this list that maps to work your team does manually right now. Not all six. One.
If you run development work, start with code-review-graph. Two commands to install. No ongoing configuration. You will see token usage change on the next session with your AI coding tool.
Run it for a week. If it saves measurable time, go deeper. If it does not, set it down and try the next one on the list.