Model releases, platform improvements and deprecation notices, posted monthly. Version numbers align with the details on each model page.
The legacy POST /v1/moderate endpoint is deprecated and will sunset on 31 March 2027. Migrate to POST /v1/moderation, which supports per-category thresholds and returns the same moderation labels with more granular confidence data.
Streaming now sustains token-level deltas under burst load, and Tamil support is available in preview alongside English, Bahasa Malaysia and Chinese. Feedback is welcome through the beta program.
Threat detection now groups related findings into campaign narratives across 90 days of telemetry. Analysts receive ranked, explainable timelines with recommended containment actions.
Low-light scene parsing is more robust and median latency drops roughly 9% across image workloads. The update is backward-compatible for existing v2.4 callers.
The content-safety model now recognises the v3 MyKad number pattern, reducing missed PII flags on Malaysian identity documents. Existing threshold policies are unaffected.
Our first OpenAI-compatible multilingual LLM is now available to all accounts. It serves English, Bahasa Malaysia, Chinese and Tamil with streaming, function calling and JSON mode.
Document extraction improves on handwritten forms, including cursive numerals in Malaysian government forms. Field-level confidence scores are now reported on handwriting-heavy documents.
Professional customers now receive 300 requests per minute, up from 150, with 20% burst headroom above the steady rate. The Developer tier remains at 60 RPM.
Inference is now available in Indonesia, joining Malaysia and Singapore. Customers can pin workloads to a single region from the console.
The vision model now produces natural-language captions in Bahasa Malaysia alongside English and Mandarin. The track covers 4K stills and video-frame analysis without changing the request schema.
Embedding quality improves for mixed EN/MS/ZH corpora, with higher nDCG@10 on cross-lingual retrieval benchmarks. The updated weights apply automatically to all callers.
The docs are reorganised around four guides with a single quickstart that reaches a first call in under five minutes. Each model page now carries full request and response schemas.