January 2026: New base model, memory architecture, and image generation engine

Aria-3 base model deployed

• Context window expanded to 32k tokens, up from 8k in Aria-2.
• Median message coherence score up 31% on internal evaluation benchmarks.
• Inference throughput up 22% at equivalent output quality.
• Persona consistency across multi-hour sessions materially improved.

Ref: ARIA-301

Long-term memory v2: semantic vector retrieval

• Replaced keyword-match storage with a semantic vector store (1536-dim embeddings).
• Retrieval accuracy up 44% vs. v1 on 30+ fact recall benchmarks.
• Storage capacity: up to 6x more facts retained per companion before eviction.
• Memory operations exposed to users via the Memories tab (view, edit, delete).

TTS model v2: 19 languages, six new voices

• Added Japanese, Korean, Italian, Dutch, Swedish, Polish, and Portuguese (BR).
• Six new voices, each trained on a minimum of 200 hours of studio audio.
• TTFA (time-to-first-audio) reduced from 1.4s to 0.9s at p50.

Video generation model v1: closed beta

• First production video generation model: lip-synced clips from a text prompt and character seed.
• Output spec: 512x512, 24fps, 10-30 second duration.
• p95 generation latency: 12 seconds on current inference hardware.
• Beta capped at 500 accounts while output quality and generation cost are validated.

Image generation: consistent-character engine v1

• Face-anchoring step added at inference time to preserve identity across a session.
• Internal face-match score across a 10-image sequence improved from 61% to 79%.
• Evaluated using a fine-tuned face-similarity model on a held-out set of 500 characters.

SOC 2 Type I completed

• SOC 2 Type I audit for the period ending December 2025 completed. Report available on request at [email protected].
• Full legal documentation at /legal-information.

Legal Information →