GLM 5.3 and Freebuff: A Fully Free AI Agent on Your Desktop

In this video Fatih Bellisoy walks through an AI agent he presents as usable without any subscription or API key, built around Z.ai's GLM 5.3. The model is framed as a computer-use system for desktop tasks, long-horizon planning and the design of complex workflows. The ability to experiment without cost pressure is the reason he now uses it in his own projects.

GLM 5.3 keeps the same base as GLM 5.2; every lift comes from scaling post-training. The team combined IndexShare for long-context efficiency, SAO for RL on long horizons and slime for large-scale asynchronous training, then grew the set of task environments collected over months. The Flash variant starts from a freshly trained base: 320 billion total parameters with only 18 billion active, a hybrid of sparse and linear attention plus Manifold-Constrained Hyper-Connections that cuts attention compute by 3x and KV cache by 4.4x at 1M context; on Z.ai Code Bench Flash jumps from 46.2 on 5.2 to 63.4, closing in on Claude Opus 4.8.

The first demo is intentionally rough: a single short prompt for a 2+1 suite hotel room in 3D that you can walk through. The result is not perfect, yet a navigable space from one sentence validates the idea of pulling visual feedback into the coding loop.

The second project is a personal, still-unfinished control center that orchestrates several agents and surfaces live work data — social metrics, view counts and latest videos in one place. It is incomplete but has been pushed forward with GLM 5.3 at no extra cost. The dashboard shows the agent can read desktop context beyond just writing code.

The third demo is a 3D chess game, again from a single prompt. Play proceeds cleanly, turns alternate and animations run without breakage. The knight model is weak on aesthetics, but shipping a rule-correct, interactive game from a terse command shows the model can handle front-end and game logic together.

You can also try the model on the web. Z.ai's chat surface offers GLM 5.3 and a faster Flash option with templates for slides and web apps. On desktop the normal route is API access with usage-based billing: Z.ai lists GLM 5.3 at $1.40 per million input tokens and $4.40 per million output tokens, with cached input at $0.26, while the Flash variant drops to $0.15 in and $0.50 out and $0.03 cached. Flash hits 57 on the Artificial Analysis Intelligence Index at $0.045 per task, delivering similar intelligence at roughly one-tenth the price. Freebuff claims to zero out that bill with ad support.

After downloading Freebuff you pick an empty folder; the agent gains read-write access to that workspace. In the model picker GLM 5.3 Flash appears as recommended. The service grants 25 tokens per day, 5 tokens buys a one-hour GLM 5.3 session. Alternatives are listed too — Mimo 2.5 at 10 tokens per hour, DeepSeek V4 Flash at 25 — and you can also attach Claude Code or Codex via your own key. The session does not start on app launch but on your first message, kicking a one-hour countdown.

For the final build planning mode is enabled to create a chat interface on a canvas where answers can be pinned as notes. The agent asks scoping questions, drafts a detailed plan and waits for approval before writing code. After revision the app lets you select a reply and add it to the canvas, which the host finds handy for language learning. Including setup and small fixes the build took about 15 minutes.