Japan’s Digital Agency placed Fujitsu Ltd’s Takane 32B large language model into the Gen-nai government generative-AI trial environment on Tuesday, letting ministries experiment with on-premise summaries of public comments filed during draft-rule consultations. Mei Kobayashi’s AI desk views the move as procurement foreplay ahead of an August target to host domestic models on Sakura Cloud infrastructure for agencies that cannot send citizen text to overseas APIs.
Gen-nai in practice
Gen-nai began as an internal chat workspace for the Digital Agency and expanded to other ministries under strict logging rules. Trials now compare Takane against NTT Data’s tsuzumi 2 and Preferred Networks’ PLaMo 2.0 Prime on identical summarization tasks: clustering repetitive objections, tagging policy themes, and drafting neutral briefs for political principals.
Human reviewers must sign off before summaries influence rule revisions; the agency’s published evaluation criteria penalize hallucinated citations.
Why Takane
Fujitsu trained Takane on Japanese legal and administrative corpora with emphasis on honorific register and nested bullet structures common in comment forms. The 32-billion-parameter size balances quality with inference cost on air-gapped servers ministries prefer for unpublished submissions.
A prior pilot at unnamed agencies reportedly cut first-pass reading time by double-digit percentages on high-volume consultations, though the agency has not published audited productivity statistics.
Security and privacy
Comment forms sometimes include personal anecdotes; on-premise inference keeps those off public clouds. The Personal Data Protection Commission’s guidelines on AI processing still require purpose limitation and deletion schedules even when models run locally.
Opposition lawmakers asked Tuesday whether vendor staff can access logs during debugging; the Digital Agency said contractors work under the same confidentiality clauses as civil servants.
Procurement timeline
Open recruitment for domestic models closed in January with 15 applicants and seven selected for trials. Paid procurement could begin in fiscal 2027 if evaluations pass accuracy and cost thresholds. Until then, ministries use trial credits without multi-year commitments.
Global vendors remain available through separate frameworks, but Gen-nai’s domestic track signals political preference for sovereign stacks.
Vendor competition
NTT Data and PFN models run on the same trial cloud, forcing Fujitsu to differentiate on administrative Japanese nuance and integration with Fujitsu’s existing government systems contracts. Errors that misstate statutory citations could disqualify Takane regardless of fluency.
Mei Kobayashi notes that model size alone does not win government work in Japan—audit trails and liability allocation do.
Evaluation metrics
Digital Agency scoring sheets weight citation accuracy higher than fluency, a change from early chatbot pilots that rewarded polished prose. Ministries must log every prompt and output hash for later audit if a summarized comment is misquoted in a cabinet decision memo.
Vendors cannot train on citizen submissions without explicit consent forms attached to each consultation portal.
Bottom line
Adding Takane 32B to Gen-nai lets ministries test sovereign summarization of public comments before a planned mid-2026 operational window, with humans still liable for every line that reaches a rulemaking file.