What This Is
A software engineer with 20 years of experience (formally trained in mechanical engineering) recently ran a comparative test on his home workstation: he put two open-source LLMs — Google's Gemma4-31B and Alibaba's Qwen3.8-27B — through two real tasks from his daily work.
The first was "understanding an unfamiliar codebase": given an open-source project he had never seen, the AI had to articulate the core abstractions, flag what callers need to watch for, and weigh the trade-offs in under 900 words. The second was "writing a project from scratch": a small C# + WPF (Windows desktop UI framework) utility that would have cost him 1-2 days of hand-coding, described in roughly 8 prompt segments, with the AI producing the code directly.
Both models were quantized to Q4 precision (quantization — using fewer bits to store each parameter so a large model fits on a consumer GPU), then exercised through two CLI runtimes — Codex CLI and OpenCode CLI — with the latter showing negligible differences.
How the Industry Sees It
Final grades: Gemma4 earned a B-, Qwen3.8 earned a B+. The engineer observed two distinct personalities. Gemma4 plays like a "fast gunslinger": precise task comprehension and fewer tool calls (~10 versus Qwen's 20-30), but the first-pass output is only barely passable and needs several feedback rounds to become shippable. Qwen3.8 plays like a "plodding student": more small mistakes (missing using statements, getting tangled in its own logic), but it aims higher and persists longer, ultimately requiring one fewer round of rework.
We think the takeaway worth highlighting is this: open-source local models have walked out of the "running demos" phase and into "delivering real work." That is genuine market pressure on closed-source vendors.
But we also need to keep our heads. This is one engineer's subjective verdict on one project, not a systematic benchmark. B+ still means multiple iteration loops before the output converges on something usable. Running a 31B quantized model locally demands an RTX 4090-class GPU, with hardware investment starting around 10,000 RMB. And against top-tier closed models like Claude and GPT, the gap on more complex tasks remains real.
What It Means for Regular People
For enterprise IT: Data never leaves the internal network, which means sensitive industries — finance, healthcare, government — can deploy local LLMs for code assistance without the data-compliance risks of cloud-based APIs.
For working professionals: Programmers can already treat local AI as a pair-programming partner, but it is nowhere close to replacing a full engineer; non-technical roles will see almost no direct use for now.
For the consumer market: Running a 31B-class model locally still requires a roughly 10,000 RMB GPU; most everyday users will reap the benefits indirectly through cloud services, not by self-hosting.