Self-Hosted Qwen and Gemma: Limits of Complex Coding

Where self-hosted models already work — and where they stall

I would not equate “the model writes code” with “the model can handle systems engineering.” In the original discussion, a developer describes a year of using in-house Qwen- and Gemma-based solutions: they are useful for routine work, but complex coding remains a weak point.

By November 2026, the question is no longer whether self-hosted coding assistants are viable in principle. The focus has shifted to performance on long-running tasks: maintaining a plan across files, understanding internal conventions, using the terminal, and recovering from failed patches without losing context.

Qwen’s materials on Qwen3-Coder describe its flagship configuration as a 480-billion-parameter MoE model with 35 billion active parameters. It was reported at about 67% on SWE-bench Verified under standard evaluation and 69.6% with up to 500 turns. That is a strong result, but it does not turn the model into a compact local tool: storing weights, supplying memory and bandwidth, and operating the stack become an infrastructure project of their own.

Gemma is easier to deploy, but has less evidence behind it for repairing large repositories. A smaller model can work well for code search, documentation, SQL, tests, and constrained refactoring. I would not extrapolate that success to kernels, distributed systems, or complex infrastructure without an internal evaluation.

There is also a less obvious boundary: a self-hosted model does not automatically mean self-hosted security. An agent still needs shell access, build tools, containers, dependencies, and credentials. A mistake in that surrounding setup can be more dangerous than a bad autocomplete.

Why model origin is not the main test for a bank

For a bank, the decisive factor is not the model’s passport but control over the entire environment. Qwen’s Chinese origin alone proves neither suitability nor unsuitability: teams still need to assess the license, provenance of weights and dependencies, data flows, jurisdictional exposure, and internal third-party policies.

A practical design is deliberately conservative: the model proposes changes in an isolated environment rather than receiving a direct route to production. It needs short-lived credentials, network and command restrictions, immutable logs, secret scanning, and mandatory human review.

My engineering conclusion is straightforward: Qwen already looks like a serious candidate for private repository work, while Gemma remains a sensible option for narrower tasks. But the real boundary is not between open and closed models. It lies between an attractive patch and a reproducibly safe system change.

We previously covered Rust LocalGPT, a single-binary local assistant designed for on-device deployment with persistent memory and an HTTP API. Its practical trade-offs help frame where self-hosted models remain viable before coding workloads become too demanding.