Preferred Networks

PFN Co-Founder and CEO Daisuke Okanohara @hillbig has been invited to speak at #ITFWorld2026, the annual forum to be held by the international nanoelectronics R&D hub Imec between May 19-20 in Antwerp, Belgium. The session theme will be announced soon! imecitf.com/world/speakers

Announcing CuPy v14! 🚀 🔹 NumPy v2 semantics 🔹 CUDA Pip Wheels support 🔹 bfloat16 & structured dtype 🔹 Enhanced NumPy/SciPy API coverage Read more on our blog for the full details! 👇

Discover unknown stable crystals at unprecedented speed. Introducing Matlantis CSP. 💎🚀 We are thrilled to announce the launch of Matlantis CSP (Crystal Structure Prediction), a new capability that rapidly identifies previously unknown stable crystal structures from the enormous search space of atomic configurations and compositions within a given elemental system. Materials R&D has long depended on repeated synthesis experiments, even when the likelihood of success is low. Matlantis CSP introduces a computational screening step earlier in the process to rule out what is physically implausible in advance. By combining Matlantis’ core technology—the universal machine learning potential PFP—with proprietary algorithms, Matlantis CSP enables you to: ✅ Screen out experimentally unviable configurations before experimentation. ✅ Explore complex composition spaces with approx. 3–6× greater efficiency than random search. ✅ Discover novel materials that do not exist in current databases. We are honored to have Honda R&D as an early adopter. “We have high expectations for CSP as a technology that will dramatically improve exploration efficiency in materials development,” said Mitsumoto Kawai, Chief Engineer, Device Process, Innovative Research Excellence, Honda R&D Co., Ltd. “Through CSP, crystal structure searches—including multi-component systems and metastable structures that were previously impractical—have become feasible. Being able to narrow down promising crystal structures and compositions with high confidence before experiments will not only increase the probability of realizing next-generation materials but also shorten development timelines.” Learn how Matlantis CSP can transform your materials discovery: 👇 matlantis.com/en/product/mtc… #Matlantis #CrystalStructurePrediction

Watch PFN co-founder @hillbig talk about AI in Japan as a guest lecturer at Columbia Business School Center on Japanese Economy and Business @ColumbiaCJEB in December last year youtube.com/watch?v=rsD9Qn…

PFN was selected as a partner for JICA's PoC for the National Bank of Cambodia🇰🇭. PFN will develop an AI model that forecasts the liquidity of the local currency to support Cambodia's Riel promotion efforts to achieve independent monetary policy. preferred.jp/en/news/pr2026…

Pre-register to our Summer 2026 Internship Program in Japan to receive invitations for applications from PFN! We welcome students who wish to take on technological challenges in algorithms, mathematics and implementation. preferred.jp/en/news/pr2026…

Here's our paper on PARC, a coding agent that can autonomously execute long-horizon tasks. PARC assesses its actions and outcomes by itself to detect and correct errors. We demonstrated its problem-solving capability in materials science and Kaggle tasks. arxiv.org/abs/2512.03549

PFN's Japanese translation-specific LLM Plamo™ Translate will be adopted by Japan's "Gennai" government AI project as a high-quality, Japan-made on-premise AI for supporting government employees. preferred.jp/en/news/pr2025…

Do you need some extra bandwidth (for LLM inference)? If you’re attending #SC25 in St. Louis, meet us at the booth 4311 and check out our MN-Core™︎ series of AI processors!

PFN will be at #SC25 in St. Louis between Nov 16-21! Meet us at the booth #4311 to learn about our MN-Core™ series of fast, powerful AI processors preferred.jp/en/news/pr2025…

What a week at #CIIE! We were thrilled to exhibit Matlantis at the L'Oréal BIG BANG BEAUTY TECH INNOVATION PROGRAM booth. A huge thank you to the L'Oréal team for this opportunity and to everyone who visited us! #CIIE #LorealBigBang

Optuna v4.6 has been released! 📊Optuna Dashboard LLM Integration 🚀Significant speedup of GPSampler 🦾Full support for multi-objective and constrained optimization in AutoSampler ⛏️Robust Bayesian Optimization Algorithms medium.com/optuna/announc…

PFN 2025 summer intern Sahoo Manish conducted a study on the automatic construction and optimization of pseudopotentials, which is crucial for accelerating DFT (Density Functional Theory) calculation. Read more: tech.preferred.jp/en/blog/automa…

We have released the latest version of our flagship LLM PLaMo 2.1 Prime with enhanced AI agent functionality. PLaMo can now automatically select tools that best suit user instructions and reference multiple databases and work with other AI agents. preferred.jp/en/news/pr2025…

Our summer intern Baoxi Huang explored Bloom filters in MinHashLSH for deduplication. Preserving accuracy while cutting 87.2% of memory use and making processing 38.4% faster on large datasets. Efficient, memory-friendly & fast! tech.preferred.jp/en/blog/improv…

AI Agenda: Japan Strikes Back As Western artificial intelligence developers accelerate international expansion, Japan, the world’s fourth-largest economy, is shaping up to be a key battleground. Read more from @JuroOsawa 👇 theinformation.com/articles/japan…

To secure the much-needed computing power of the next generation of our MN-Core AI processors, we are testing high-density servers with a direct liquid cooling system that efficiently manage the heat from the chips. preferred.jp/en/news/pr2025…

We are accepting pre-entries from students expected to graduate in 2027 for new hires. Entrants will receive direct notices from PFN. Please register here: preferred.jp/en/news/2027-n…

Optuna v4.5 has been released! ⛏️GPSampler for constrained multi-objective optimization 🚀Significant speedup of TPESampler and plot_hypervolume_history 🦾CmaEsSampler now supports 1D search space 🐍The optunahub library is published on conda-forge github.com/optuna/optuna/…

We have developed a caching system for container images in #kubernetes. By introducing a new caching tier within a cluster, we have achieved faster image pulls and reduced network traffic. Blog post: tech.preferred.jp/en/blog/kubern…