HY Image 3.5 Preview Review: Pricing, Text, and Editing

👁 112 reads

Tencent released HY Image 3.5 Preview on OpenRouter on October 5, 2026, priced at $0.024 per generation. We ran it through two real-world tasks from our AI Image Generator Arena: a multi-step photo edit executed flawlessly across three turns, and a stress test on Polish typography where it correctly rendered 10 out of 13 target texts. Unlike its predecessor, this release is closed-weights.

Tencent HY Image 3.5 Specs and Features

HY Image 3.5 Preview is a unified model designed for text-to-image generation, reference-based synthesis, and conversational image editing where users iteratively modify an existing image. According to Tencent:

• +30% win rate in human evaluations compared to HY Image 3.0.
• In Tencent’s blind test involving hundreds of designers, it reportedly matched Seedream 5.0 Pro and slightly edged out Nano Banana Pro and Qwen-Image-3.0 Pro. This is Tencent’s internal benchmark, and independent verification is still pending.
• Key gains focus on in-image typography, layout fidelity, photorealism, and visual consistency across multi-turn edits.
• Output resolutions scale up to 2K; our test generations produced 1536 x 1536 px (square) and 1856 x 1248 px (horizontal edit).

One critical change: its predecessor, HY Image 3.0 (80B MoE architecture with 13B active parameters), offered open weights. For version 3.5, Tencent has not disclosed the architecture, parameter count, or any plans for open weights. You cannot run this model locally on your own hardware. For self-hosted alternatives, check our ranking of open-source image models.

Test 1: Non-English Text Rendering

We tested it in Polish to gauge how well it handles non-English orthography and complex diacritics. We used the same four prompt challenges as in our standard benchmark: a promotional retail poster, a traditional bistro menu board, a book cover heavy with diacritics, and a dumpling package label - totaling 13 distinct text strings to generate.

Where it stumbled:

• Retail Poster: It printed „paźziernika” instead of „października” (“October”) in the bottom row, while the main headline and discount line „-40% NA WSZYSTKO” (“-40% off everything”) were spelled perfectly.
• Bistro Menu: It produced „GOLĄBKI” missing the crossed L (should be „GOŁĄBKI”), and printed the currency marker lowercase instead of uppercase. Other dishes („Żurek”, „pierogi ruskie”, „kompot”, and „szarlotka”) came out cleanly.
• Product Packaging: It misspelled mushrooms as „Pierogi z kapustą i grzydami” instead of „grzybami”, and hallucinated an unprompted phrase, „SUCHE MIESZANKA DO MROŻONEK” (“dry mix for frozen goods”).

Counting manually, 10 out of 13 text elements were completely accurate. Automated line-by-line OCR scrapers gave it only 3/13 because the model broke text across unexpected line wraps. For context under that same strict line parser, GPT Image 2.5 Sunburst scored 7/13, FLUX 2 Max reached 6/13, Nano Banana 2.1 hit 5/13, and Ideogram 4.5 scored 4/13. The takeaway: letterforms and diacritics are sharp, but it can still scramble unfamiliar vocabulary like most image generators. Any graphic meant for publication still requires careful proofreading.

Test 2: 3-Step Iterative Image Editing

Using an original photograph of a street-level bakery, we ran a sequential edit where each step builds directly on the previous generation: repaint the front doors bottle green, place an „OTWARTE 7-22” (“OPEN 7-22”) vinyl sticker on the glass, and swap the storefront sign to read „PIEKARNIA ZOSIA” (“ZOSIA BAKERY”).

• All three requested edits were executed precisely, text rendered cleanly, and the replacement sign preserved the original font style and color scheme.
• Zero background drift: a bystander, a parked bicycle, a sidewalk chalkboard reading „CHLEB 6,90 zł”, a red car, and the cobblestone street remained entirely unchanged.
• Latency ranged from 26 to 44 seconds per step, billed at $0.024 each.

While three turns is too brief to formally enter our arena leaderboard (which demands 10 consecutive edits alongside 4K detail retention), this trial demonstrates that targeted conversational editing is arguably the model’s strongest asset.

Pricing and Where to Access HY Image 3.5

A technical note for developers: on OpenRouter, this is an image-only model and will not respond via the standard chat/completions route. You must query the dedicated /api/v1/images endpoint and supply reference images for editing inside the input_references array. For comparison, Nano Banana 2.1 costs approximately $0.034 per 1K image via API (details in our Nano Banana guide).

Who Is It For?

• Recommended: Product photo touch-ups and localized “edit this specific detail” workflows - fast, inexpensive, and resistant to scene drift.
• Use with caution: Complex multilingual signage or typography - character shapes are solid, but spelling must be checked manually.
• Not recommended: Workflows requiring self-hosted, local GPU inference - HY Image 3.5 does not provide open weights.

For side-by-side evaluations across other commercial tools, explore our comprehensive guide to AI image generators.

// read moreStepFun Step 5 Preview Hits OpenRouter: Kimi K3 Rival at $1NewsGrok Imagine Image 2.0: Spot Editing, Sharp Text & 5 InputsAI ToolsAI Image Generator Arena: Compare 11 Models Side-by-SideBenchmarks

Kinetyra