Articles tagged "Japanese LLM"

**Preferred Networks (PFN) officially released PLaMo 3.0 Prime on June 22, 2026** ([official press pr20260622](https://www.preferred.jp/ja/news/pr20260622) / [tech blog](https://tech.preferred.jp/ja/blog/plamo-3-0-prime-release/)). Successor to PLaMo 2.0 Prime (2025 Nikkei Excellent Product Award grand prize), this is the production rollout after a 3-month monitor program following the March 19, 2026 beta. **Context extended from 64K beta → 256K production**, **dual Reasoning / Non-reasoning variants**, proprietary tokenizer optimized for Japanese token efficiency, post-training with SFT + DPO + RL. **Compared against** gpt-oss-120b / Qwen3.6-27B (open) and GPT-5.4 mini / Claude Haiku 4.5 (closed in the same price tier). **Evaluated on 15 benchmarks**: JFBench / IFBench / Japanese MT-Bench / lawqa_jp / MedRECT / Japanese Medical Licensing Exam / MT-Bench / AIME 2024 / GPQA-Diamond / BFCL / LongBench v2 / HELM Safety. PFN CEO/CTO Daisuke Okanohara claims parity or superiority over same-tier models on Japanese instruction-following, coding, and tool-use, while ITmedia at the beta stage noted weakness in math and multi-tool selection. **Pricing is aggressive**: Standard plan **¥60 input / ¥250 output per 1M tokens** (up to 128K), Free plan pending, Provider plan custom-quoted. **Distribution**: PLaMo API (SaaS), on-premise, Amazon Bedrock Marketplace, Snowflake. **Prime itself is closed-weights**, but NICT-co-developed base models `plamo-3-nict-2b/8b/31b-base` are open on Hugging Face. **Adoption**: standard model in miibo / Tachyon / QommonsAI, and **selected as a trial model for the Japanese Digital Agency's common generative-AI platform 'Gennai'**. **Not disclosed**: parameter count, dense vs MoE, and independent third-party benchmark verification — pending Nejumi LLM Leaderboard registration.

On-device deep dive on Liquid AI's two Japanese-specialized models released on Hugging Face in early June 2026: **`LFM2.5-1.2B-JP-202606`** (language, 1.17B params, 32K context) and **`LFM2.5-Audio-1.5B-JP`** (1.2B language core + 115M FastConformer encoder, 24kHz, Speech-to-Speech). Grounded in the official model cards and Liquid AI blog. Covers the **Liquid Neural Network-derived architecture** (16 layers = 10 LIV convolution + 6 GQA), the sub-2B-class leadership benchmarks (JMMLU 54.19 / J-MIFEval 79.08 / J-GSM8K 62.20), audio ASR with CommonVoice 8 (ja) CER 4.42 — about half of Whisper-large-v3 — but **trailing Whisper on JSUT and ReazonSpeech (domain-dependent gap)**, the non-Apache **LFM Open License v1.0**, hardware support across Apple Silicon / AMD Ryzen AI / Qualcomm / NVIDIA / mobile CPU, competitive positioning vs Gemma 4 12B / Qwen 3.5 / TinySwallow / Sarashina, and adoption guidance for Japanese on-device AI, call centers, meeting notes, and in-person retail.