Six tools dropped in the week of September 20. Every one of them is free or open-source. This is what each one does, which ones pair well, and which one is worth installing first.
Number 6: OpenViking Gives Your Agents a Shared Memory
Every AI agent you run today forgets everything the moment a session ends. You paste context manually. The agent makes the same mistake twice. The same setup question gets answered every time.
OpenViking is a context database. Your agent reads from it at the start of a task and writes back when the task finishes. The next agent, or the next session of the same one, starts where the last left off.
It is open-source, built specifically for multi-agent workflows. One database, any number of agents, memory that accumulates instead of resetting.
Why it ranks sixth and not first: memory only compounds value if you already have agents running. If your setup is still mostly manual, install this once you have two agents sharing work. The compound effect kicks in fast after that.
Where to get it: search OpenViking on GitHub. The README covers setup and the API is straightforward.
Number 5: Google Open-Sourced ARTEMIS. It Turns Prompts into Real Phone Actions.
ARTEMIS is a Google open-source project. You write a prompt. ARTEMIS executes it as a real automation on a phone screen: no app, no shortcut scripting, no workflow builder.
It connects over MCP and integrates directly into Claude Code, Codex, or Antigravity setups. While it runs, it captures logs and screenshots so you can verify what happened and where it failed.
The use case: anything that lives on a phone screen and has no API. Customer service apps. Multi-step mobile forms. Approval workflows locked inside apps you cannot touch from a server.
The honest limit: ARTEMIS runs on a physical device, so high-volume automation is not its use case. For triggered, one-off phone actions, it is the cleanest tool available.
Where to get it: Google released it publicly. Find it under the Google org on GitHub.
Number 4: PrismML's Ternary Bonsai 2 Runs a 27B Reasoning Model on Your Own Machine
Ternary Bonsai 2 is a 27-billion-parameter reasoning model. PrismML uses Ternary compression to shrink the weights enough to run locally without a GPU on most setups.
What it handles: coding, math, tool calling, image understanding. Context window: 262,000 tokens. That context length is longer than most cloud models charge for at the top tier.
The practical question: is a 27B local reasoning model good enough to replace a cloud call for your most common tasks? For classification, summarization, and document review, the answer is usually yes.
The honest limit: it will not match a frontier 200B+ model on complex multi-step reasoning. Local models trade capability ceiling for privacy and zero per-call cost. That trade is worth making for the right tasks.
Where to get it: PrismML released it on Hugging Face. Search Ternary Bonsai 2.
Number 3: AWS Open-Sourced Pizza Bot. It Is an Inbox for Background Agents.
Background agents have a visibility problem. They run, and you have no clean way to see what finished, what failed, or what needs your decision. You either build a monitoring layer yourself or you fly blind.
Pizza Bot is AWS's answer to that. It is an inbox for background agents. Everything an agent completes, everything that failed and needs a retry, everything that needs a human decision: it surfaces in a unified view formatted like email.
It runs locally on Mac, Windows, and Linux. No cloud dependency.
Why this is worth installing now: the further agents run from human review, the more a missing review layer costs. A single bad agent run that goes unnoticed is more expensive than the time it takes to set this up.
Where to get it: AWS released it on GitHub. Search Pizza Bot AWS.
Number 2: GLM-5.3 FlashX. 320 Billion Parameters. Free on Your Own Hardware.
Zai released GLM-5.3 FlashX: 320 billion parameters, natively multimodal, one-million-token context, 200 tokens per second, free to download and run.
That context window changes the economics of large-document tasks. You can feed entire codebases, months of contracts, or years of email into a single call. No chunking, no summarization pipeline, no context management overhead.
At 200 tokens per second locally, it is not slower than a cloud API with network round-trips. At zero marginal cost per call, it shifts the math on any high-volume task.
The catch: 320 billion parameters requires real hardware. Expect two or more high-end GPUs. This is not a laptop model. If you have a server with that capacity, the economics justify evaluating it seriously.
Where to get it: Zai released GLM-5.3 FlashX on Hugging Face.
Number 1: Unsloth Studio. Fine-Tune 500-Plus Models in a Free Colab. No Code.
Fine-tuning has required Python skills, clean labeled datasets, GPU setup, and hours of debugging. Unsloth Studio removes all of that.
Pick a model from over 500 available. Upload your dataset. Hit train. Export and run locally.
No code. No cloud bill. The Colab notebook handles the infrastructure.
Why this ranks first: every other release this week is a tool you add to what you already do. Fine-tuning is a capability upgrade to every model you already use. A model fine-tuned on your own data follows your instructions more reliably, maintains your tone, and makes fewer errors on your specific tasks than a base model with the same prompt.
One training run on your support tickets, your client emails, or your internal documentation produces a model that outperforms the base version on those tasks. That improvement does not require better prompts. It just persists.
Where to get it: Unsloth Studio is available on Google Colab. Search Unsloth Studio fine-tuning.
Which One to Install This Week
Start with Unsloth Studio if you have any dataset larger than 100 labeled examples: customer emails, product descriptions, support transcripts. One fine-tuning run is worth more long-term than any other single tool here.
Add Pizza Bot second if you are running background agents without a review layer. The missing review layer is a liability that compounds quietly.
OpenViking third, once you have two agents that share work across sessions. Persistent memory turns a series of isolated runs into a compounding system.
ARTEMIS, Ternary Bonsai 2, and GLM-5.3 FlashX are worth watching. Install them when a task maps directly to what they do.
What These Releases Tell You
Every tool this week is free. Every one is open-source. The cost of running a capable multi-agent setup is collapsing.
The constraint is not access. It is the time to evaluate, configure, and integrate. That constraint does not go away on its own.
The previous week's roundup covered six releases that leaned heavily on code intelligence. This week's releases lean on infrastructure: memory, monitoring, fine-tuning. They are not standalone tools. They are the connective tissue between the tools you already run.
The Mistakes That Make These Fail
Fine-tuning requires a clean dataset. If your data is inconsistent or unlabeled, the model will learn those inconsistencies. Fix the dataset before you blame the model.
Pizza Bot surfaces what went wrong after the fact. It does not prevent bad agent behavior. Design your agents to fail gracefully; Pizza Bot is the review layer, not the error layer.
GLM-5.3 FlashX requires hardware most small teams do not have. The benchmark numbers are real. Hardware requirements gate them.
OpenViking compounds value only if agents are already writing useful information back to it. A memory store fills with noise if agents write carelessly. Define what goes in the store before you deploy it.
If You Only Do One Thing This Week
Run one fine-tuning job in Unsloth Studio. Pick your most repetitive AI task, gather 100 to 500 examples of good output, and train. If the first run does not improve results, check the dataset before adjusting anything else. The dataset is almost always the problem.