【ローカルLLM】国産LLM「LLM-jp-4-33B」と「Qwen3.8-27B」の推論性能比較（RTX 5070 Ti + RTX 3070 Ti）A hands-on benchmark comparing the newly released Japanese LLM "LLM-jp-4 33B"…

A new hands-on benchmark has put Japan's latest open large language model, "LLM-jp-4 33B," head to head with Alibaba's Qwen3.8-27B, running both models locally on a consumer dual-GPU rig built from an RTX 5070 Ti and an RTX 3070 Ti. The test matters because it addresses a practical question facing many practitioners: whether a newly released domestic Japanese model can hold its own against a well-regarded international competitor when both are run on hardware that enthusiasts and small teams can actually afford, rather than on data-center accelerators.

According to the article, the LLM development group behind the project (referred to as LLMC) released LLM-jp-4 33B in the week before publication. The model is a Dense architecture with roughly 33.2 billion parameters, and it is described as the Dense counterpart to a Mixture-of-Experts (MoE) model the same group published in April of this year. That pairing is notable: releasing both Dense and MoE variants lets users choose between the two design philosophies depending on their hardware and latency needs, and it gives the community a cleaner basis for comparison since the two share lineage.

The comparison target, Qwen3.8-27B, comes from Alibaba's widely used Qwen family, which has become a common reference point in local-LLM circles thanks to its strong multilingual performance and permissive availability. Pitting a 33B Dense Japanese model against a 27B competitor is a reasonable matchup on paper, though the difference in parameter counts and architectures means results should be read as indicative rather than definitive. The author also notes that LLMC has previously issued a statement regarding performance comparison and verification, which appears to frame how such benchmarks should be interpreted, though the specifics of that statement are not fully reproduced here.

The hardware choice is central to the exercise. Combining an RTX 5070 Ti with an RTX 3070 Ti pools VRAM across two cards of different generations, a pragmatic setup for hobbyists who upgrade incrementally rather than buying matched pairs. Running a 33B-parameter Dense model typically requires either substantial memory or quantization to fit within consumer GPU limits, so the effective performance depends heavily on the quantization level, context length, and the inference backend used. These variables tend to influence tokens-per-second throughput and output quality as much as the underlying model weights do.

For readers less familiar with the ecosystem, local inference of this kind is usually handled by tools such as llama.cpp, Ollama, LM Studio, or vLLM, each offering different trade-offs between ease of use, quantization support, and multi-GPU handling. Quantization methods like GGUF variants or AWQ compress model weights to reduce memory footprint, which is often the deciding factor in whether a large model runs at all on a given card. Splitting a model across two GPUs, as done here, can unlock larger models but may introduce communication overhead that affects speed, so real-world benchmarks like this one are valuable complements to headline accuracy scores.

The distinction between Dense and MoE architectures is worth keeping in mind as well. Dense models activate all parameters for every token, offering predictable behavior and often simpler deployment, while MoE models route each token through a subset of "expert" networks, which can improve efficiency at scale but complicates memory planning. LLMC providing both forms means users can weigh consistency against efficiency for their particular workload.

The broader context is a period of rapid activity in sovereign and language-specific LLM efforts. Japan's LLM-jp initiative has positioned itself as an openly documented, research-driven program aimed at strengthening Japanese-language capability, an area where general-purpose international models have sometimes lagged. Benchmarks like this one, published on community platforms such as Qiita, help fill the gap between official model cards and everyday deployment realities.

As with any single hands-on test, the findings are best treated as a data point rather than a verdict. Differences in prompts, quantization, driver versions, and backend configuration can meaningfully shift outcomes, and the author's setup reflects one specific configuration. Still, the comparison offers useful signal on how a fresh domestic Dense model behaves against an established rival on accessible hardware.