Japanese and Korean AI models: LLM-jp, K-EXAONE and the GPUs they need

Part of our AI models guide: 11 posts on choosing and running AI models.
The best Japanese-built model tested, NII's LLM-jp-4.1 33B, ranks about 100th on the Nejumi Japanese leaderboard, with US and Chinese models at the top. Korea's open flagships score 8 to 9 on Artificial Analysis's mostly English v5 index, against 46 for the best open model. There is still plenty to download: Japan's LLM-jp-4.1, Swallow and Rakuten AI 3.0, and Korea's K-EXAONE 2.0 (LG), A.X K2 (SK Telecom), Solar Open 2 (Upstage) and Motif 3. PLaMo 3.1 Prime, Sarashina3, tsuzumi 2, Sakana Namazu and Solar Pro 4 are API or paid-licence only. Several "sovereign" models are foreign open models tuned locally: Rakuten AI 3.0 is built on DeepSeek-V3, Swallow on OpenAI's gpt-oss and Alibaba's Qwen3, and Namazu on Moonshot AI's Kimi K2.6. Most Japanese open models fit on one or two GPUs; Rakuten AI 3.0 and Korea's two largest take an 8-GPU H200 server, though A.X K2 is validated on four B300 GPUs.
Japanese AI models at a glance
Open weights are model files you can download and run yourself, which is not the same as open source (see open-weight vs closed models). B is billion parameters; a token is the smallest unit of text an AI model reads and writes, often part of a word; context is how many tokens a model reads at once. A mixture-of-experts (MoE) model uses only its "active" parameters per token but must hold them all in GPU memory.
Model | Maker | Size | Licence | Can you download it? | Built on |
|---|---|---|---|---|---|
LLM-jp-4 and 4.1 | National Institute of Informatics (NII) | 8B and 33B dense; 32B total, about 3.8B active (MoE, named 32B-A3B) | Apache 2.0 | Yes | Own model, 11.7 trillion training tokens (model card) |
GPT-OSS Swallow; Qwen3 Swallow | Institute of Science Tokyo and AIST | 20B and 120B; 8B, 30B-A3B and 32B | Apache 2.0 | Yes (20 Feb 2026) | OpenAI's gpt-oss; Alibaba's Qwen3 |
Rakuten AI 3.0 | Rakuten | 671B total, 37B active (MoE) | Apache 2.0 | Yes (3.0.1 posted 4 Nov 2026) | DeepSeek-V3-0324, per its model card |
ELYZA-Thinking-1.0 | ELYZA (KDDI group) | 33B; 32B total, about 3.8B active (32B-A3B) | Apache 2.0 | Yes (2 Oct 2026) | NII's LLM-jp-4 |
PLaMo 3 (base and Translate) | Preferred Networks (base models with NICT) | 0.8B to 31B | PLaMo Community Licence | Yes, after a licence click-through (Translate since 5 Nov 2026) | Own model |
PLaMo 3.1 Prime | Preferred Networks | Not disclosed | Commercial | No: API or on-premises (since 27 Oct 2026) | Own model |
Sarashina3 | SB Intuitions (SoftBank) | Not disclosed | Commercial | No: API, free beta since 13 Oct 2026 | Own model; built from scratch, per SB Intuitions |
tsuzumi 2 | NTT | About 30B (press reports) | Commercial | No: on-premises or private cloud; one GPU, per NTT | Own model; built from scratch, per NTT |
Sakana Namazu; Sakana Fugu | Sakana AI | Not disclosed | Commercial | No: via Sakana's API; Fugu in Ultra and Max tiers | Namazu: Moonshot AI's Kimi K2.6. Fugu: routes work across an unnamed pool of models |
NEC's cotomi v3 and Fujitsu's Takane (built on Cohere's Command R+) are closed models sold to companies that want data kept in Japan. Stockmark-2-100B is another self-trained open model.
Two traps. PLaMo's community licence allows free commercial use only after registration, and only for companies with annual revenue up to 1 billion yen. And Rakuten AI 3.0 uses 37B of its 671B parameters per token, but all 671B must sit in GPU memory. For US and Chinese models, see how many GPUs you need to run an LLM.
Korean AI models at a glance
Model | Maker | Size | Licence | Can you download it? | Worth knowing |
|---|---|---|---|---|---|
K-EXAONE 2.0 | LG AI Research | 750B total, 37B active (MoE); 262K context | Apache 2.0 | Yes (31 Jul 2026) | 10 languages, including Korean and Japanese |
K-EXAONE 2.5 | LG AI Research | 750B, per LG | Not announced | No: free API since 9 Nov 2026 | Further-trained K-EXAONE 2.0 |
EXAONE 4.0 32B; EXAONE 4.5 33B | LG AI Research | 32B; 33B | EXAONE licence 1.2 NC | Yes, research and education only | Commercial use needs a separate licence from LG |
A.X K2 | SK Telecom | 688B total, 33B active (MoE); 256K context | Apache 2.0 | Yes (29 Jul 2026) | Trained from scratch, per its model card |
Solar Open 2 | Upstage | 250B total, 15B active (MoE) | Upstage Solar Licence | Yes (22 Jul 2026) | English, Korean and Japanese |
Solar Pro 4; Solar Mini 4 | Upstage | Pro: not disclosed; Mini: 35B total, 3B active | Commercial | No: via Upstage's API, or on-premises by request | 512K context, per Upstage |
Motif 3 | Motif Technologies | 314B total, 13.2B active (MoE) | MIT | Yes (Aug 2026) | English and Korean; cut from the national project in Aug 2026 |
Korea's flagships are large open-weight MoE models, built for the science ministry's contest. LG says it upcycled K-EXAONE 2.0, starting from its own earlier K-EXAONE rather than from blank weights. K-EXAONE 2.0, A.X K2 and Motif 3 allow commercial self-hosting; so does Solar Open 2, but derivatives must carry the Solar name and a "Built with Solar" credit. Smaller open models, Kakao's Kanana-2 and Naver's HyperCLOVA X SEED, need a separate licence above 10 million monthly users; Kanana also needs one to resell access by API or cloud, and HyperCLOVA X SEED for a product that competes with Naver's own. Licence summary only; not legal advice.
How they score against global models
Global models still lead on independent tests. For Japanese, the Nejumi Japanese leaderboard (Nejumi Leaderboard 4), snapshot of 22 October 2026 as compiled by Qualiteg:
Model | Made in | Open weights? | Score | Rank |
|---|---|---|---|---|
Claude Opus 5 (Anthropic) | US | No | 0.872 | 1st |
Qwen3.8-2.4T-A95B (Alibaba) | China | Yes | 0.860 | 3rd (top open model) |
LLM-jp-4.1 33B (NII) | Japan | Yes | 0.721 | About 100th (best Japanese-built model tested) |
Closed Japanese flagships (PLaMo 3.1 Prime, Sarashina3, tsuzumi 2) are not tested there. For Korea, no independent Korean-language leaderboard covers the top models; this is Artificial Analysis's general, mostly English Intelligence Index v5 (19 November 2026):
Model | Maker | Weights | Score |
|---|---|---|---|
Claude Fable 5.2 (max) | Anthropic, US | Closed | 65 (top) |
Kimi K3.1 Pro | Moonshot AI, China | Open | 46 (best open model) |
Solar Pro 4 | Upstage, Korea | Closed | 14 (peer median 22) |
A.X K2 | SK Telecom, Korea | Open | 9 (open peer median 23) |
K-EXAONE 2.0 | LG AI Research, Korea | Open | 8 (open peer median 23) |
In August 2026, on its earlier index version, Artificial Analysis called Korea a clear number three in AI after the US and China.
On their own model cards (vendor-reported), Korean makers compare against Chinese open models and small US ones, never top closed US models. KMMLU-Pro covers Korean professional exams; CLIcK, Korean culture and language:
Model card | Korean model's scores | Global models on the same card |
|---|---|---|
A.X K2 (SK Telecom) | KMMLU-Pro 80.5; CLIcK 91.6 | DeepSeek V4-Flash 78.8 and 89.6; Qwen3.5 78.4 and 88.4 |
Solar Open 2 (Upstage) | KMMLU-Pro 78.4; CLIcK 90.7 | DeepSeek V4-Flash 78.9 and 89.2; GPT-5.4 mini (closed) 78.1 and 89.6 |
K-EXAONE 2.0 (LG AI Research) | KMMLU-Pro 69.1; CLIcK 84.2 | DeepSeek V4-Pro 80.5 and 91.6; Qwen3.5 77.4 and 88.9 |
Tuning has a trade-off: Rakuten's card (vendor-reported) shows Rakuten AI 3.0 beating its DeepSeek base on Japanese culture questions (JamC-QA 76.9% against 70.2%) but losing on maths (MATH-100 86.9% against 90.9%). Local models win on local knowledge; test both on your own prompts.
GPU requirements for each open model
Weights take the parameter count times the bytes per parameter: 2 at BF16 (16-bit), 1 at FP8 (8-bit), about half at FP4 (4-bit). Add room for the KV cache, the model's working memory for each conversation, which grows with context (see our GPU memory guide). NVIDIA's NVFP4 format is built for Blackwell GPUs (B200, B300), which have native FP4 maths. Per GPU: H100 80 GB, H200 141 GB, B200 180 GB, B300 288 GB; an 8-GPU server holds 640 GB, 1,128 GB, 1,440 GB or 2,304 GB.
Model | Memory for the weights | Smallest GPU count that holds it (sizing only) | Source | What one 8-GPU server gives you |
|---|---|---|---|---|
LLM-jp-4.1 33B, Qwen3 Swallow 32B, PLaMo 3 31B, Kanana-2 30B-A3B | About 61 to 66 GB at BF16 | One H200, or one H100 (tight) | FlexiCloud estimate | Eight copies on an 8x H200 or H100 server, or room to fine-tune |
GPT-OSS Swallow 120B | About 234 GB at BF16 | Three H200 or two B200; the 4-bit MXFP4 version fits one 80 GB GPU, like gpt-oss-120b | FlexiCloud estimate | Several copies; the card advises contexts up to 32K tokens |
Rakuten AI 3.0 (671B, MoE) | About 675 GB (FP8 weights) | One 8x H200 server | Model card shows 8 GPUs; H200 is a FlexiCloud estimate | One copy on the whole server |
K-EXAONE 2.0 (750B, MoE) | About 750 GB at FP8; about 1.5 TB at BF16 | One 8x H200 server at FP8 | Model card | One copy; B200 and B300 servers leave more room for long contexts |
A.X K2 (688B, MoE) | About 695 GB at FP8; plan for 860 GB or more (card: 647 and 800 GiB) | Four B300 (validated); the NVFP4 version on four B200 | Model card | One copy on half an 8x B300 server; 8x H200 also holds it (FlexiCloud estimate) |
Solar Open 2 (250B, MoE) | About 500 GB at BF16 | Four H200 | Model card: minimum 4x H200, recommended 8x H200 | The recommended setup |
Motif 3 (314B, MoE) | About 630 GB at BF16; about half at FP8 | Two B200 with the NVFP4 version | Model cards: NVFP4 on 2x B200; BF16 and FP8 tested on B200 and H200, with 8-GPU examples | One copy at FP8, or several NVFP4 copies on a B200 server |
Further training an open model on Japanese or Korean text, as for Swallow, Rakuten AI 3.0 and Namazu | Depends on the base model | Several 8-GPU B300 servers or a GB300 NVL72 rack | FlexiCloud guidance | Ask for cluster or rack terms |
FlexiCloud arranges whole 8-GPU servers (nodes); the GPU-count column is sizing only. Tell us the model, context length, languages and traffic, and we will size the node.
Japanese and Korean use more tokens
APIs bill per token (see our token pricing guide). Total tokens for the same FLORES-200 text, compared with English (Petrov et al., NeurIPS 2023 data; ratios calculated by FlexiCloud):
Tokenizer | Japanese | Korean |
|---|---|---|
GPT-4 (cl100k) | 2.30x | 2.38x |
Original Qwen | 1.46x | 1.64x |
Llama 1 and 2 | 2.24x | 3.18x |
On the same measure, GPT-4's tokenizer needs 4.79x for Hindi and 3.04x for Arabic (see our India, Gulf and Europe posts). An English chatbot's 100 million tokens a month become about 238 million in Korean on GPT-4's tokenizer. Larger multilingual vocabularies narrow the gap, as the Qwen row shows, and LLM-jp-4 and A.X K2 ship their own tokenizers, per their cards. On your own GPUs, extra tokens mean fewer answers per GPU-hour; test throughput in your language.
National programmes and the NVIDIA deal
Japan funds model builders and pushes its own models into government; its second AI Basic Plan (14 July 2026) seeks "open AI sovereignty": no heavy dependence on any one country or company. Korea runs a knockout contest and buys GPUs for the state.
Country | Programme | What it is | Status on 19 Nov 2026 |
|---|---|---|---|
Japan | GENIAC (METI and NEDO, since Feb 2024) | Pays part of the computing cost for foundation-model builders | Round 4 chosen from 50 applications (4 Jun 2026): domain-specific models, 6 to 12 months each |
Japan | Noetra | National foundation-model company; SoftBank, Sony, NEC and Honda at the core, 44 investors; Preferred Networks and AIST take part | Selected 30 Jun 2026; 387.3 billion yen for FY2026; reportedly up to about 1 trillion yen over five years, reviewed yearly |
Japan | Government AI "GENAI" (Digital Agency) | Rolling out to about 180,000 central-government staff, with a push for Japanese-built models | Five Japanese models under contract for trials; blind A/B tests Sep to Nov 2026; paid procurement planned from FY2027 |
Korea | Sovereign AI foundation model project (Ministry of Science and ICT) | Teams must train their own models from blank weights; fine-tunes of foreign models do not qualify | Five teams in Aug 2025, three left: SK Telecom, Upstage, LG AI Research; two finalists due about mid-Feb 2027 |
Korea | NVIDIA supply announcement (31 Oct 2025) | More than 260,000 GPUs over several years: up to 50,000 for the government, the rest for Korean industrial groups | Unchanged, the government said on 13 Nov 2026, after NVIDIA's change of chief executive |
Korea | Government GPU purchases | About 13,000 B200 and H200 GPUs; first batch delivered Dec 2025 | First support round: 13,712 GPUs requested, 4,224 allocated |
Data rules in Japan and Korea
Japan. The APPI, Japan's privacy law, needs prior consent naming the country before personal data goes abroad, unless the country is approved (the EU, the UK) or the recipient meets equivalent standards. The AI Promotion Act (2025) has no penalties.
Korea. PIPA, Korea's privacy law, needs consent for overseas transfers unless an exception applies, such as an approved destination (the EU since September 2025); since 11 September 2026, fines can reach 10% of total revenue in specified cases. Under the AI Basic Act (in force since 22 January 2026), foreign AI firms over set thresholds must name a local representative, with fines up to 30 million won after at least a year's grace.
Public-sector pushes such as GENAI and Korea's "AI for All" chatbot (at least 50% on sovereign-standard Korean models) cover public services, not private buyers. Self-hosting open weights keeps prompts away from a model maker's API; choose the region by where personal data must stay (see bare metal vs cloud GPUs). Not legal advice.
Japanese and Korean models: picking a server
Most Japanese open models need one or two GPUs, while Korea's two largest take an 8-GPU H200 server. Every server we arrange is a whole 8-GPU machine, from H100 to B300, so one can carry multiple copies of a small model; for further training, ask for a GB300 NVL72 rack. Each server comes from a verified operator in one of our six regions (North America, the UK, Europe, India, Asia Pacific and the Gulf), for twelve months or longer, billed upfront, and setting up the model software stays with you.
When you know the model, languages, traffic and where personal data must stay, start a sizing request; our GPU servers page lists the hardware.
Where these figures come from
Checked 19 November 2026 unless dated.
- Hugging Face cards: LLM-jp-4.1, GPT-OSS Swallow (MXFP4 version), Rakuten AI 3.0, ELYZA, PLaMo 3, K-EXAONE 2.0, A.X K2 (NVFP4), Solar Open 2, Motif 3 (NVFP4); licences: EXAONE 4.0, Kanana-2, HyperCLOVA X SEED.
- Impress Watch on PLaMo 3.1 Prime (27 Oct 2026) and PLaMo 3 Translate (5 Nov 2026); Sarashina3 (29 Oct 2026); NTT (20 Oct 2025); Sakana Namazu and Fugu; Fujitsu (16 Jul 2024); LG changelog (9 Nov 2026); Upstage Solar Pro 4 and Mini 4.
- Scores: Qualiteg (22 Oct 2026); Artificial Analysis. Tokens: Petrov et al.
- Programmes: NEDO on GENIAC; Honda on Noetra (16 Jul 2026); Jiji Press (30 Jun 2026); Digital Agency; AI Basic Plan (14 Jul 2026); NVIDIA (31 Oct 2025); Korea Times (1 Dec 2025); Digital Today (20 Jan 2026, GPU support); ZDNet Korea (3 Nov 2026); Yonhap via Newstong (5 Oct 2026); Herald (13 Nov 2026).
- Data rules: DLA Piper on Japan and Korea; FPF; Cooley (27 Jan 2026); Asia Economy (13 Jul 2026).