Qubitz and the local-agent fever: when data sovereignty outweighs raw intelligence
🕒 Published on Zendoric: July 3, 2026 · 01:20
An independent developer releases Qubitz, an AI agent that runs entirely locally on open models from 7B to 35B, with no cloud or subscriptions. It's a tiny project, but it captures well where a growing part of the open-weight ecosystem is pushing.
By GitHub (a project by Gabrieliam42) · July 2, 2026.
Qubitz is a standalone AI agent designed to run entirely on the user's computer, with no cloud calls or subscriptions. It runs on GGUF models served with llama.cpp —from Qwen 3.5 9B to 27B and 35B variants such as GLM 4.7 Flash, Gemma 4 31B, GPT-OSS 20B or a MoE called Ornith-1.0-35B— and adds local context retrieval via embeddings (BAAI/bge-code-v1), a GUI and CLI interface, and an MCP server mode to connect to other Tools-compatible utilities. Its design is oriented toward WSL2/Windows environments, with an explicit bridge for running hybrid workspaces.
What's interesting is not the model's sophistication —the author himself acknowledges in the README that Qubitz doesn't compete in raw intelligence with frontier cloud agents— but the architecture: a 'harness' and a wrapper that control routing, task execution and tool access, leaving the small model (7B-35B) with only the language and reasoning it actually needs to make decisions. It's a direct response to a known problem with mid-sized local models: without an external control layer, they tend to ignore instructions, misinterpret tools or drift off task. Wrapping those weaknesses in deterministic logic, instead of entrusting everything to the model, is the only reasonable way to get an 8B or a 30B to behave predictably on real repository tasks.
We should be honest about the proportions: this is a repository with two stars and one fork, no comments on Hacker News and a single karma point—that is, a personal project at a very early stage, not a launch with market traction. There are no verifiable performance figures here, no comparison against cloud agents, and no evidence of real adoption beyond the author himself. Any reading of its technical quality must remain a design promise, not a demonstrated fact.
That said, as sector context, Qubitz is a small but representative example of a trend that is real and that we have been pointing to: the open frontier (Qwen, GLM, Gemma, GPT-OSS, DeepSeek) has come down enough in size and cost that any developer with a 12-24 GB GPU can build a functional agent without depending on OpenAI, Anthropic or Google. That move toward local—driven by privacy, control and the absence of recurring billing—is exactly the facet of AI democratization that interests us most in the long run: not all the value of future abundance will come from the big labs, but also from thousands of modest projects that make it viable to have useful artificial intelligence without handing over your data or your budget to a provider. That most of these projects go nowhere doesn't invalidate the direction; it simply reminds us that the road to abundance is made of many small experiments, and only a few will survive to become real infrastructure.