Qwen3.5-9B（Q4/6.6GB）にM1 Maxで日本語を書かせたら、答えは131字なのに出力は3936トークンだったHands-on testing of Qwen3.5-9B (Q4, 6.6 GB) on an M1 Max revealed that a…

A recent hands-on write-up published on Zenn takes aim at a claim that circulates frequently in local large language model communities: that Qwen3.5-9B, running at Q4 quantization in just under 6 GB of memory, can match or even surpass GPT-4 on Japanese-language tasks. The author set out to test this on real hardware—an Apple Silicon M1 Max with 64 GB of unified memory—using a locally stored copy of qwen3.5:latest weighing 6.6 GB. The exercise matters because "beats GPT-4" taglines are common in this space, rarely verified under real conditions, and easy to accept at face value.

The headline outcome is a caveat rather than a confirmation. The author states plainly that whether the model "beats GPT-4" could not be measured. A credible verdict would require human evaluation across a large set of Japanese prompts, and the comparison is further complicated by the practical difficulty of accessing GPT-4 under identical conditions. What the test could measure was something more concrete: what actually happens when you run this model on an M1 Max.

Here the numbers are striking. A Japanese answer that amounted to only 131 characters consumed 3,936 tokens to produce. That gap between a short final response and the volume of tokens generated exposes what the author frames as a significant verbosity issue. In practical terms, the model appears to spend a large amount of computation and output on intermediate text before arriving at a brief conclusion, which directly affects latency and throughput on local hardware regardless of how "smart" the final answer may be.

Some context helps explain why this can happen. Many recent models, including those in the Qwen family, support reasoning or "thinking" behavior in which the model generates extended intermediate steps before producing a concise answer. When that mode is active, the token count for a task can balloon far beyond the length of the visible response. The 131-character-to-3,936-token disparity is consistent with that pattern, though the write-up focuses on the observed measurement rather than attributing a single definitive cause.

The setup also illustrates several prerequisite concepts worth understanding. Q4 quantization compresses a model's weights to roughly four bits of precision, shrinking its memory footprint so it can fit in a few gigabytes and run on consumer machines. That is what allows a 9-billion-parameter model to occupy about 6.6 GB. Apple Silicon is well suited to this because its unified memory architecture lets the GPU address the same large memory pool as the CPU; on a 64 GB M1 Max, memory capacity is not the binding constraint, so the practical limits are more about speed and output efficiency than about whether the model loads at all.

The tooling around this ecosystem is relevant background too. The qwen3.5:latest naming convention is characteristic of Ollama, a popular front end that packages models for local use and sits atop inference engines such as llama.cpp. On Macs, Apple's MLX framework offers an alternative optimized path for running quantized models. These tools have made it dramatically easier to download and run capable models offline, which is precisely why community claims about their quality spread quickly and why independent, reproducible measurements remain valuable.

Japanese adds its own wrinkle to benchmarking. Tokenization does not map neatly to characters, and Japanese text can be tokenized less efficiently than English depending on the vocabulary, so token counts and per-character costs are not directly comparable across languages. Standardized Japanese evaluation suites such as JGLUE exist for more rigorous scoring, but they still cannot resolve the GPT-4 comparison the way informal claims imply, since the two systems run under different conditions and access constraints.

The broader takeaway is a measured one. The article does not declare Qwen3.5-9B good or bad, nor does it dismiss the "beats GPT-4" assertion outright; it simply demonstrates that the claim is difficult to verify in a real local environment and that verbosity is a concrete, observable cost. For readers weighing local models, the lesson is likely that quantization and memory fit are only part of the picture. Output efficiency, reasoning behavior, and reproducible measurement matter just as much as headline capability claims, and they are far easier to check on your own machine.