BRIEF · OPEN-SOURCE MODELS · GLOBAL COMPARISON · jdd-kami
Snapshot as of June 18, 2026. Covers China, USA, France, Germany, Japan, Korea, India, and the EU. Sources: BenchLM, Presenc AI, klymentiev.com, lilting.ch, NII, arXiv, etc. Model development moves fast — rankings shift weekly.
Field notes: "Scale" = total parameters (e.g., 1.6T = 1.6 trillion); "Active" = active parameters per inference (MoE architecture); "Context" = max context length (e.g., 1M = 1 million tokens). Data as of 2026-06-18.
| Country | Lab | Flagship Model | Scale | Active | License | Strengths | Context |
|---|---|---|---|---|---|---|---|
| China | DeepSeek DeepSeek |
DeepSeek V4 Pro | 1.6T | 49B | MIT | Coding SWE-Bench 83.7%, lowest cost | 1M |
| China | Alibaba (Qwen) Alibaba |
Qwen 3.5 397B | 397B | 17B | Apache 2.0 | All-rounder, multilingual, multimodal | 256K |
| China | Moonshot (Kimi) Moonshot AI |
Kimi K2.6 | ~1T | ~32B | MIT | Coding HumanEval 99% | 200K |
| China | Z.ai (Zhipu) Zhipu AI |
GLM-5.2 | 744B | 40B | Open | Knowledge MMLU 96, BenchLM #1 | 128K |
| China | Tencent (Hunyuan) Tencent |
Hunyuan 3.0 | 295–406B | MoE | Open | Multimodal: text, video, 3D world | 128K |
| China | Huawei (Pangu) Huawei |
Pangu Ultra MoE | 718B | — | Open | Trained from scratch on Chinese hardware stack (Ascend) | 128K |
| China | ByteDance (Doubao) ByteDance |
Seed 2.0 | Multi-tier | — | Proprietary | Lowest price $0.022/M output | — |
| China | Baidu (ERNIE) Baidu |
ERNIE 5.0 | — | — | Mixed | LMArena top-10, 1.5B daily API calls | — |
| China | 01.AI (Yi) 01.AI |
Yi-Lightning | Mixed | — | Mixed | Pioneer, Yi-Coder still strong | — |
| China | MiniMax MiniMax |
abab series | Multi-tier | — | Proprietary | Long-context multimodal, HKEX listed | — |
| USA | Meta Meta AI |
Llama 4 Maverick | 400B | 17B | Llama Community | Largest ecosystem, 1M context | 1M |
| USA | Meta Meta AI |
Llama 4 Scout | 109B | 17B | Llama Community | 10M ultra-long context (longest open-source) | 10M |
| USA | Google Google |
Gemma 4 medium | ~27B | dense | Gemma | Single-GPU inference, safety-tuned | 256K |
| USA | OpenAI OpenAI |
GPT-oss 120B | 120B | dense | Apache 2.0 | OpenAI-compatible, drop-in replacement | 128K |
| France | Mistral AI Mistral AI |
Mistral Medium 3.5 | 675B | 41B | Apache 2.0 | EU-friendly, GDPR compliant, enterprise support | 128K |
| France | kyutai Non-profit |
Moshi / Helium | Small-scale | — | Open | Voice model Moshi | — |
| France | Pleias Pleias |
Pleias 1.0 | — | — | Public domain | Strictly open data | — |
| Germany | Aleph Alpha Aleph Alpha |
Pharia-1 / Luminous | — | — | Proprietary-leaning | Enterprise sovereignty, auditable, on-premise | — |
| EU | Silo AI (AMD) Finland |
Poro / Viking | — | — | Open | Multilingual EU, OpenEuroLLM co-lead | — |
| EU | OpenEuroLLM 20+ organizations |
In development | — | — | Open | Covers all EU languages | — |
| Japan | NII National Institute of Informatics |
LLM-jp-4-32B-A3B | 32B | 3.8B | Apache 2.0 | Trained from scratch (11.7T tokens), MT-Bench JA 7.82 | — |
| Japan | NVIDIA (JP) NVIDIA Japan |
Nemotron Nano 9B JP | 9B | — | NVIDIA Open | Sovereign AI, first sub-10B | — |
| Japan | Tokyo Tech / AIST Tokyo Tech |
Swallow 30B-A3B | 30B | 3B | Open | Continued pretraining + RL, strong OCR correction | — |
| Japan | Preferred Networks PFN |
PLaMo 2.0 | 31B | — | Not disclosed | Trained from scratch | — |
| Japan | NEC NEC |
cotomi v3 | — | — | Not disclosed | Trained from scratch | — |
| Japan | Sakana AI Sakana AI |
Namazu | Multi-scale | — | Depends on base | Post-trained on DeepSeek | — |
| Korea | Upstage Upstage |
Solar Open | 102B | MoE | Solar-Apache-2.0 | Bilingual (Korean-English) | 128K |
| Korea | Naver Naver |
HyperCLOVA X Seed 32B | 32B | — | LimitedOpen | Leading downloads in Korea | — |
| India | Sarvam AI Sarvam |
Sarvam 105B / 30B | 105B | MoE | Open | India's largestOpenmodel、Trained from scratch、22+ Indian languages | — |
| Rank | Model | BenchLM Score | Strengths | License |
|---|---|---|---|---|
| #1 | GLM-5.2 Z.ai（Zhipu AI） |
94 | Knowledge, reasoning, agentic | Open |
| #2 | Qwen 3.7-Max Alibaba（Alibaba） |
90 | Coding, full-context | Apache 2.0 |
| #3 | Qwen 3.7-Plus Alibaba（Alibaba） |
88 | Balanced | Mixed |
| #4 | DeepSeek V4 Pro DeepSeek（DeepSeek） |
87 | Coding (strongest open-source), low cost | MIT |
| #5 | Kimi K2.6 Moonshot（Moonshot AI） |
~85 | Coding (HumanEval 99%) | MIT |
10 labs at or near frontier, vs only 3–4 in the West (Meta, Mistral, AI2). DeepSeek is the only fully MIT top model.
Z.ai (HKEX Jan) and MiniMax both went IPO, shifting capital structure and disclosure cadence.
Only lab with both frontier models and frontier training chips (Ascend). Unique hardware sovereignty angle.
Trained from scratch（LLM-jp-4）vs Continued pretraining（Swallow、Nemotron JP）vs Post-training（Namazu）。Three paths, three sovereignty strategies。
Not competing on scale but on auditability, GDPR compliance, multilingual coverage. Mistral is the anchor lab ($13.7B valuation).
Sarvam AI is the firstTrained from scratch 105B -class model，focusing on 22+ Indian languages。
MostChinalabs open base layers、close flagship layers。DeepSeek（fully MIT）is the exception。OpenOpen-source has caught up with closed-source frontier on most benchmarks。
Most labs shipped 4–7 flagship updates in the past 12 months. Model development moves fast — rankings shift weekly.
DeepSeek, Alibaba (Qwen), Moonshot (Kimi), Z.ai (GLM), Tencent (Hunyuan), Huawei (Pangu), ByteDance (Doubao), Baidu (ERNIE), 01.AI (Yi), MiniMax. Top 5 BenchLM scores range from 94 to ~85 — densest globally.
Meta (Llama 4 Maverick / Scout) — largest ecosystem, 10M context. Google (Gemma 4) — single-GPU inference. OpenAI (GPT-oss 120B) — OpenAI-compatible.
Mistral AI（$13.7B valuation、Apache 2.0、GDPR compliance）。Aleph Alpha（Enterprise sovereignty、auditable）。kyutai（Non-profit、Voice）。Pleias（Public domain）。Silo AI（Multilingual EU）。OpenEuroLLM consortium（20+ organizations、Covers all EU languages）。
Trained from scratch：NII LLM-jp-4（11.7T tokens、Apache 2.0）、PLaMo 2.0、cotomi v3。Continued pretraining：Nemotron Nano 9B JP（NVIDIA）、Swallow 30B-A3B（Tokyo Tech/AIST）。Post-training：Namazu（Sakana AI、on DeepSeek ）。
Upstage Solar Open（102B MoE、Solar-Apache-2.0、Korean-English bilingual）。Naver HyperCLOVA X Seed 32B（Leading downloads in Korea）。
Sarvam AI（Sarvam 105B / 30B）——India's largest Openmodel，Trained from scratch，focusing on 22+ Indian languages。2026 3 Open。