Japan

AI models, vendors, and hardware buying realities from Japan. Curated by myaihardware.com.

29
Models
6
Vendors
2
Languages

Why Japan matters for AI builders

Japan is a critical node for AI builders because it is one of the few places where high-bandwidth memory (HBM) and advanced packaging substrates are still manufactured, making it a bottleneck for GPU supply chains. Its dominant model families, like the open-source LLM series from Stability AI Japan and local fine-tuned variants of Llama and Qwen, reflect a strong preference for efficiency and domain-specific adaptation over raw scale. An underrated angle is that Japan's aging industrial and robotic infrastructure creates a massive early-adopter demand for on-device AI that must run on legacy hardware, forcing builders to optimize for limited memory and thermal constraints before the rest of the world catches up.

Market context

METI has allocated significant funding to accelerate domestic large language model development as part of Japan's sovereign AI push, aiming to reduce reliance on foreign models and secure strategic autonomy. The government's AI Strategy 2024-2025 prioritizes compute infrastructure and evaluation benchmarks while confronting the intrinsic challenge of Japanese tokenization, where complex script mixing and morphological ambiguity inflate token counts and degrade efficiency. This funding directly targets companies and research consortia building models with native script-aware tokenizers to overcome the linguistic bottlenecks that handcuff generic architectures.

Models (29 of 29)

Filter by vendor, license, or parameter range.

Stockmark-13b

Stockmark Inc. · Japan

Open
Params
13B
Context
4,096
Min VRAM
8 GB
GGUF
No

Excellent Japanese fluency, strong instruction following.

Region slug: japan

Stockmark-100b

Stockmark Inc. · Japan

Restricted
Params
100B
Context
8,192
Min VRAM
50 GB
GGUF
No

Massive knowledge capacity, state-of-the-art Japanese generation.

Region slug: japan

karakuri-lm-70b-chat-v0.1

Karakuri AI · Japan

Open
Params
70B
Context
4,096
Min VRAM
35 GB
GGUF
Yes

Fluent conversational Japanese, good adherence to prompts.

Region slug: japan

karakuri-lm-8x7b-chat-v0.1

Karakuri AI · Japan

Open
Params
8B
Context
32,768
Min VRAM
8 GB
GGUF
Yes

Long context window, efficient inference comparable to 13B dense model.

Region slug: japan

Swallow-7b-hf

Tokyo Institute of Technology / NII · Japan

Open
Params
7B
Context
4,096
Min VRAM
4 GB
GGUF
Yes

Lightweight, strong Japanese base model.

Region slug: japan

Swallow-13b-hf

Tokyo Institute of Technology / NII · Japan

Open
Params
13B
Context
4,096
Min VRAM
8 GB
GGUF
Yes

Balanced performance, good for fine-tuning.

Region slug: japan

Swallow-70b-hf

Tokyo Institute of Technology / NII · Japan

Open
Params
70B
Context
4,096
Min VRAM
35 GB
GGUF
Yes

Highest Japanese capability among Swallow versions, strong zero-shot.

Region slug: japan

Swallow-MS-7b-v0.1

Tokyo Institute of Technology / NII · Japan

Open
Params
7B
Context
8,192
Min VRAM
4 GB
GGUF
Yes

Very fast inference, permissive license, good Japanese.

Region slug: japan

Swallow-MX-8x7b-NVE-v0.1

Tokyo Institute of Technology / NII · Japan

Open
Params
8B
Context
32,768
Min VRAM
10 GB
GGUF
Yes

Long context, efficient MoE, strong Japanese performance.

Region slug: japan

Llama-3-Swallow-70B-v0.1

Tokyo Institute of Technology / NII · Japan

Open
Params
70B
Context
8,192
Min VRAM
35 GB
GGUF
Yes

Superior Japanese understanding with latest Llama 3 architecture.

Region slug: japan

ELYZA-japanese-Llama-2-13b

ELYZA Inc. · Japan

Open
Params
13B
Context
4,096
Min VRAM
8 GB
GGUF
Yes

Very high Japanese chat quality, widely used in Japan.

Region slug: japan

ELYZA-japanese-Llama-3-8B

ELYZA Inc. · Japan

Open
Params
8B
Context
8,192
Min VRAM
4 GB
GGUF
Yes

Fast, lightweight, excellent Japanese for its size.

Region slug: japan

PLaMo-13B

Preferred Networks · Japan

Restricted
Params
13B
Context
2,048
Min VRAM
8 GB
GGUF
No

Trained from scratch on high-quality Japanese data, solid base model.

Region slug: japan

PLaMo-100B

Preferred Networks · Japan

Restricted
Params
100B
Context
4,096
Min VRAM
50 GB
GGUF
No

Exceptional knowledge and reasoning, near GPT-4 on Japanese.

Region slug: japan

EvoLLM-JP-v1-7B

Sakana AI · Japan

Open
Params
7B
Context
4,096
Min VRAM
14 GB
GGUF
No

High Japanese performance via evolutionary model merging, cost-effective 7B model.

Region slug: japan

EvoVLM-JP-v1-7B

Sakana AI · Japan

Open
Params
7B
Context
4,096
Min VRAM
14 GB
GGUF
No

Japanese multimodal capabilities, text and image understanding.

Region slug: japan

Sarashina2-7B

SB Intuitions · Japan

Restricted
Params
7B
Context
4,096
Min VRAM
14 GB
GGUF
No

Excellent Japanese language model, cost-effective 7B.

Region slug: japan

Sarashina2-13B

SB Intuitions · Japan

Restricted
Params
13B
Context
4,096
Min VRAM
26 GB
GGUF
No

Larger size improves reasoning and Japanese fluency.

Region slug: japan

Sarashina2-70B

SB Intuitions · Japan

Restricted
Params
70B
Context
4,096
Min VRAM
140 GB
GGUF
No

Top-tier Japanese LLM performance.

Region slug: japan

Fugaku-LLM-13B

Fujitsu · Japan

Open
Params
13B
Context
4,096
Min VRAM
26 GB
GGUF
No

Trained on Fugaku supercomputer, strong Japanese capabilities.

Region slug: japan

RakutenAI-7B

Rakuten · Japan

Open
Params
7B
Context
4,096
Min VRAM
14 GB
GGUF
No

Commercial-friendly Apache 2.0, strong Japanese performance.

Region slug: japan

RakutenAI-7B-chat

Rakuten · Japan

Open
Params
7B
Context
4,096
Min VRAM
14 GB
GGUF
No

Good at conversational Japanese tasks.

Region slug: japan

NTT tsuzumi (7B)

NTT · Japan

Restricted
Params
7B
Context
4,096
Min VRAM
14 GB
GGUF
No

Efficient, good performance in Japanese, designed for enterprise use.

Region slug: japan

CyberAgentLM2-7B

CyberAgent · Japan

Open
Params
7B
Context
4,096
Min VRAM
14 GB
GGUF
No

Open source, strong Japanese generation.

Region slug: japan

calm3-22b-chat

CyberAgent · Japan

Open
Params
22B
Context
4,096
Min VRAM
44 GB
GGUF
No

Largest open Japanese chat model from CyberAgent, good instruction following.

Region slug: japan

OpenCALM-7B

CyberAgent · Japan

Open
Params
7B
Context
2,048
Min VRAM
14 GB
GGUF
No

Pioneering open Japanese LLM, good base.

Region slug: japan

llm-jp-13b-v2.0

LLM-jp (NII, etc.) · Japan

Open
Params
13B
Context
4,096
Min VRAM
26 GB
GGUF
No

Open source, strong Japanese benchmarks, community-driven.

Region slug: japan

llm-jp-172b-instruct

LLM-jp (NII, etc.) · Japan

Open
Params
172B
Context
4,096
Min VRAM
344 GB
GGUF
No

Largest open Japanese LLM, strong performance.

Region slug: japan

nekomata-14b

rinna · Japan

Open
Params
14B
Context
32,768
Min VRAM
28 GB
GGUF
No

Based on Qwen2.5, strong Japanese due to continuous pretraining and fine-tuning.

Region slug: japan

Notable vendors

Tokyo Tech SwallowELYZAPFN PLaMoSakana AISB IntuitionsLLM-jp

Run benchmarks

llama.cpp, MLPerf, and inference leaderboards.

VRAM calculator

Will this model fit on your GPU?

Hardware guides

Local LLM setup, homelab, mini-PC builds.

Methodology

How we benchmark, verify ASINs, and source data.

Beginner's guide

New to local AI? Start here.

Sourcing & commerce disclosure

Model metadata on this page is compiled from each vendor's public HuggingFace repo, model card, or research paper at time of indexing. Performance figures (KMMLU, MMLU, HAERAE, AraBench, IndicGEM, etc.) link directly to source benchmarks where available; missing or partial values are shown as ", ".

Outbound hardware links (GPUs, mini-PCs, accelerators) are Amazon Associate links using store ID fredoline-20. As an Amazon Associate we earn from qualifying purchases. Prices and availability shown elsewhere on the site are snapshots, not live quotes — verify on Amazon before purchasing. See About / disclosures for the full policy.