Tencent has released a preview of Hy Image 3.5, a hosted image-generation model that merges text-to-image and image-to-image workflows under a single model ID. The preview is served through Tencent Cloud's API. It also appears in the Miora design agent and the OnSolo platform, where access is free for two weeks.

The company claims a 30% human-evaluation win-rate improvement over Hy Image 3.0, but that figure comes from internal testing with no shared benchmark. The model is listed as hy-image-v3.5-preview. Generated images are priced at $0.024 each for 2K output, and uploaded reference images are billed at no charge.

Confirmed

  • Model ID and access: hy-image-v3.5-preview on Tencent Cloud, with synchronous API responses — no job submission and polling required
  • Modes: text-to-image, image-to-image, and multi-turn editing with conversation history carried across requests
  • Output and references: documentation cites up to 4,096 × 4,096 pixels (4K) and up to 14 reference images per request, while promotional materials cite up to 2K and up to five references
  • List price: $0.024 per generated image (international) / ¥0.15 per 2K image (China); reference images free
  • Integrations and base model: Yuanbao chatbot, film-editing and design tools, Miora and OnSolo; built on the open-source 80B-parameter MoE Hunyuan Image 3.0 base (64 experts, 13B active parameters)

Unknown

  • Independent replication of the claimed 30% win-rate gain over Hy Image 3.0 — no public benchmark set, rater instructions, or sample size disclosed
  • 4K pricing and regional rate-card confirmation — the $0.024 figure is explicitly for 2K output
  • SLA, latency guarantees, concurrency limits, and data-retention terms for the preview tier
  • Stable GA release path and timeline — the model is explicitly labeled a preview
  • Exact reference-image limit per endpoint (5 vs. 14 cited in different materials)

Our take

The interesting bet here is consolidation: one model handling generation and multi-turn editing could simplify tooling that today stitches together separate models, and the conversation-history carry-over is what makes that real rather than cosmetic. That said, the headline 30% improvement is Tencent grading its own homework — internal designer evaluations against Seedream 5.0 Pro, Nano Banana Pro, and Qwen-Image-3.0 Pro are directional at best. Until independent benchmarks land and the preview's gaps (4K pricing, regional availability, stability guarantees) are filled, Hy Image 3.5 looks like a promising direction rather than a proven successor to 3.0.

Sources