Qwen 35Bの品質を7つの質問で採点したら、GPT-4に勝てるのは3領域だけだったA hands-on benchmark pitting locally-run Qwen 35B against GPT-4 across seven…

Local large language models have long been dismissed as hobbyist toys, but a hands-on benchmark posted on Zenn suggests that assumption deserves a second look. The author ran Qwen 35B locally on a single consumer-grade RTX 4070 graphics card and scored it against GPT-4 across seven questions, concluding that the open model wins outright in exactly three domains.

The framing matters because it inverts a common piece of received wisdom in the field: that a "smarter," larger model accessed through an API is broadly superior for nearly any task. The author admits to having held that view, leaving the RTX 4070 idle for roughly three months on the assumption that a 35B-class model running at home would be comprehensively beaten by a hosted model like GPT-4. Only after assembling what the post calls a "winning configuration" did the actual measurement begin.

The result, presented as a win-loss map across the seven prompts, is more nuanced than a clean sweep in either direction. GPT-4 appears to retain its edge across most of the tested categories, consistent with expectations for a large hosted model. But the author reports that Qwen 35B clearly lands in three specific areas, indicating that local inference is not simply a degraded version of an API call. The takeaway is less about raw capability and more about fit: certain practical use cases favor a model you run yourself, even when it is smaller and cheaper to operate.

Some technical context helps explain why this is notable. An RTX 4070 ships with 12GB of video memory, which is not enough to hold a 35B-parameter model at full precision. Running such a model on that hardware typically depends on quantization, where weights are compressed to lower bit depths to shrink the memory footprint, often at some cost to accuracy. That the author was able to get a 35B-class model running and competitive at all on mid-range consumer hardware is part of the story, and the "winning configuration" language suggests careful tuning of the runtime, quantization level, and prompts.

Qwen is the open model family developed by Alibaba, and it has become one of the more widely used options for people running models on their own machines, alongside families such as Meta's Llama, Mistral, and Google's Gemma. The broader local-LLM ecosystem has matured quickly, with tools like Ollama, llama.cpp, and LM Studio lowering the barrier to downloading, quantizing, and serving models without cloud dependencies. Those tools are a large part of why experiments like this one are now feasible for individuals rather than only well-resourced labs.

The appeal of local models generally rests on factors beyond benchmark scores. Running inference on hardware you own means data never leaves the machine, which is attractive for privacy-sensitive or regulated work. It also removes per-token API costs and rate limits, which can matter for repetitive or high-volume tasks. Those advantages are structural rather than measured in a seven-question test, but they help explain why finding even a few domains of clear parity or superiority is meaningful to practitioners weighing local versus hosted options.

Readers should treat the specifics with appropriate caution. Seven questions is a small sample, the scoring is a single author's hands-on judgment rather than a standardized public benchmark, and results are likely sensitive to the exact model build, quantization, and prompts used. The post itself frames the exercise as a practical shared scorecard rather than a definitive ranking, and the author notes that a separate, distinct article covering related material was published on Dev.to.

Even with those caveats, the report is a useful data point in an ongoing shift. As open models improve and consumer GPUs grow more capable, the question is increasingly not whether local LLMs can match a frontier API everywhere, but where specifically they are good enough or better. Identifying three such domains, rather than none, is the kind of incremental evidence that is gradually changing how developers decide which tool to reach for.