SB Intuitions released open weights Tuesday for a Japanese-language summarizer tuned on publicly archived House of Representatives committee transcripts, giving municipal clerks a locally hosted option as Ministry of Internal Affairs and Communications logging rules for generative AI expand in October. The model targets staff who must digest multi-hour hearings on tax offsets and defense supplements without sending audio to overseas APIs.
Model scope
The weights package, branded TSUBASA-Committee-7B, ingests timestamped Japanese text—not raw audio—and outputs bullet summaries with speaker tags when diarization metadata is present. SB Intuitions trained on National Diet Library releases through August 2026, excluding closed executive sessions. It does not attempt statutory interpretation; disclaimers warn users to verify quotes against official Hansard PDFs.
Performance slides published with the release show ROUGE scores competitive with general-purpose 13-billion-parameter models on committee jargon, with lower hallucination rates on proper nouns like ministry names and bill numbers. The trade-off is narrow domain focus: sports and entertainment transcripts degrade sharply, by design.
Deployment posture
SB Intuitions ships Docker images for on-prem GPU servers common in prefectural IT closets—NVIDIA L4 class or better. SoftBank Corp enterprise customers can buy managed hosting, but the open weights allow cities with data-sovereignty policies to air-gap inference. Digital Agency procurement templates updated in September list committee summarization as an allowed pilot use when logs stay inside government VLANs.
MIC’s October rules require retention of prompts and outputs for certain public-sector AI tools. SB Intuitions bundles a logging sidecar that writes JSON audit files compatible with agency schemas; municipalities must still provide storage and backup.
Who might adopt
Osaka and Yokohama legislative liaison offices told InfoHandle they will trial the model for briefing mayors before extraordinary Diet votes on consumption-tax grocery offsets. Smaller towns with one policy staffer see the biggest gain: a three-hour welfare committee hearing collapses into five pages before breakfast briefings.
Opposition researchers warned that summarizers can flatten dissenting amendments; SB Intuitions recommends feeding only official transcript text, not live caption streams that may contain errors. National Diet Library said it does not endorse third-party models but welcomes tools that link back to canonical PDF page numbers.
Competitive context
Preferred Networks and Fujitsu already market on-prem summarizers for manufacturers; SB Intuitions differentiates on Diet vocabulary and open weights. Overseas vendors face procurement hurdles when data might leave Japan, especially for security-related committees discussing standoff missiles and cyber budgets.
Risks and limits
Open weights mean bad actors can fine-tune for misinformation, though doing so for live Diet speech would still require fake audio distribution—a separate crime. SB Intuitions withheld training code but published evaluation sets so academics can probe bias across parties. Models struggle with archaic legal kanji compounds rare in modern news; human review remains mandatory for external press releases.
Practical rollout
Clerks should pin model version hashes in meeting minutes when summaries inform decisions. Update GPUs to driver versions listed in SB Intuitions’ release notes before October 5 Diet opening. If summaries feed chatbots on municipal websites, add visible timestamps and links to source PDFs—MIC guidance treats unsourced AI text as misinformation risk during election seasons.