February 2026: Streaming inference, voice call infrastructure, and video generation v1.1

Streaming inference pipeline deployed

• TTFT (time-to-first-token) reduced from 3.2s to 0.8s across all platforms.
• Tokens streamed incrementally as generated; no wait for full response to complete.
• Achieved by switching from batched to streaming decode with a revised KV-cache layout.
• p99 TTFT is 1.4s under normal load.

Ref: ARIA-312

Live voice call infrastructure: beta

• WebRTC-based voice pipeline with sub-400ms round-trip latency on stable connections.
• Opus codec at 24kbps with VAD-gated transmission to reduce bandwidth.
• Adaptive bitrate on degraded networks (bitrate ladder incomplete at launch; see known issues).

Video generation v1.1: 720p output

• Resolution increased from 512x512 to 1280x720.
• Temporal consistency score improved from 0.71 to 0.79 on internal motion-stability benchmark.
• p95 generation time increased to 18s at 720p vs. 12s at 512px (hardware-bound).
• Closed beta expanded from 500 to 2,000 accounts.

Image generation: prompt adherence improvements

• Fine-tuned the image generation model on a curated set of 40k character-scene pairs.
• CLIP similarity score (prompt-to-image) improved by 0.06 on the internal eval set.
• Consistent-character engine applies to all new image requests automatically.

iOS app general availability

• Full App Store release; TestFlight beta closed February 28.
• Mobile inference client updated: reduced cold-start latency by 340ms on iPhone 13+.

Age verification in 8 regions

• Age verification integrated via a third-party identity vendor across 8 regions.
• Vendor listed in /legal-information.

Legal Information →