Tencent releases Hy Image 3.5 Preview at the exact moment Alibaba opens Apsara: the Chinese image model war shifts into high gear

Tencent drops Hy Image 3.5 Preview at the exact moment Alibaba opens Apsara: China's image model war shifts into high gear

🔎 A launch calibrated like a surgical strike

On September 22, 2026, Tencent put Hy Image 3.5 Preview online. That very same day, Alibaba opened its Apsara conference, its flagship annual AI event. This is no calendar coincidence: it's a tactic, and Bloomberg via The Edge Malaysia says so explicitly — the launch landed on "the exact day" of Apsara's opening.

Image generation has become the hottest front in China's AI race. ByteDance is pushing hard with Seedream, Alibaba is rolling out Qwen-Image, and Tencent just made it clear it had no intention of watching from the sidelines. The players are no longer content to simply respond to one another: they now deliberately time their moves to each other's events to steal the spotlight.

In a single day, we got both of the moment's competing strategies: a closed model at a bargain price on Tencent's side, and a spectacular open-source release with a locked-down license on Alibaba's side. We sifted the announcements through primary sources — API docs, press releases, financial media. Here's what really matters, especially if you're a developer.

Key Takeaways

• Aggressive timing: Hy Image 3.5 Preview was released on September 22, 2026, the exact day of the opening of Alibaba's Apsara Conference.
• Marketing figure: +30% capabilities vs Hunyuan Image 3.0, measured by internal blind GSB tests. No independent evaluation to date.
• Aggressive API: Chat/Messages protocol, multi-turn editing, up to 20 reference images, 4K possible, and 0.15 RMB per 2K image (September 2026, verify on Tencent Cloud) — billing only for generated images.
• On Alibaba's side: Qwen-Image-2.1 (7B) and two 9B prompt-rewriting checkpoints released as open weights… but the license moves away from Apache 2.0 to strictly non-commercial use.
• The market validated the buzz: Tencent's stock jumped sharply on the announcement, according to MarketWatch.

Recommended Tools

Timing calibrated to the very day: the synchronized-fire tactic

Tencent deliberately chose the opening day of Apsara to launch Hy Image 3.5 Preview — and it's not the first time it has pulled a maneuver like this. In China, the announcement date has become a weapon in its own right: GLM-5.3-Flash and Qwen3.8-Flash-Next were already released on the same day, a sign that this synchronization has become a reflex.

The logic is simple and devastatingly effective. When Alibaba gathers the press and developers for its annual conference, the country's entire tech attention is already mobilized. Releasing your own model on that day means capturing a share of the media buzz without spending a single yuan on PR. The market followed suit: Tencent's stock climbed sharply in the hours that followed, according to MarketWatch.

On the other side, Alibaba didn't show up empty-handed. At Apsara, Eddie Wu announced a future flagship model with 5 to 10 trillion parameters as well as an in-house accelerator chip, the Zhenwu V900, according to The Arabian Post. The Qwen3.6 family on the LLM side already shows that Alibaba is playing on every front — language, images, silicon.

One last signal not to be overlooked: talent. Tian Yonglong, who came from OpenAI where he worked on vision, joined the Hunyuan team back in July. When a Chinese giant recruits a vision specialist trained at its arch-rival just weeks before a major launch, that's not a minor HR detail — it's an admission of ambition.

My take: this unapologetic timing says something healthy about the maturity of the Chinese market. We've moved into duel-style announcements, almost media boxing. For us — observers and developers — that's excellent news: competitive pressure speeds everything up, prices included.

Hy Image 3.5 Preview: what are Tencent's "+30%" claims really worth?

Let's be direct: the "+30% overall capabilities" figure put forward by Tencent comes from blind comparative tests (GSB methodology) conducted with hundreds of designers… all internal to Tencent. It's a PR number, not an independent benchmark.

The GSB methodology (Good/Same/Bad — better, equal, worse) involves having evaluators compare two images without knowing where they came from. It's an honest tool for measuring progress between two versions of the same model. It's a dangerous tool for comparing two different vendors, especially when the evaluators work for one of them.

On this internal basis, Tencent claims that Hy Image 3.5 Preview is on par with ByteDance's Seedream 5.0 Pro and sits slightly ahead of Google's Nano Banana Pro and Alibaba's Qwen-Image-3.0 Pro. Superpower Daily makes this clear: these claims remain unquantified and independently unverified. Give it the benefit of the doubt, but verify anyway.

What the model can actually do, capability-wise

Where things get interesting is the functional scope, which is factual and verifiable in the docs:

• Text-to-image and image-to-image, with multiple image ratios.
• Multi-turn conversational creation: the model keeps the history, so you iterate as if in a chat.
• Up to 5 reference images per request on the product side, and up to 20 via the API (≤ 20 MB per image).
• Synchronous output: no async jobs or polling, the response comes back directly.

The product positioning is revealing too. The Hunyuan press release relayed by GMT EIGHT talks about moving "from beautiful images to professional aesthetic scenarios": commercial posters, storyboards, audiovisual materials, UI design, illustration, brand visuals, e-commerce. In other words: Tencent is no longer targeting the amateur general public — it's targeting your creative teams.

The model is already integrated into Yuanbao, ima, WorkRally, OnSolo, Miora and WorkBuddy, according to TechFlow. This isn't a showcase model: it's an ecosystem building block, backed by Hy3, the MoE language model released in July and already deployed in Yuanbao, WorkBuddy and Tencent Cloud.

The API Under the Hood: Messages, 4K, and 0.15 RMB per Image

For developers, this is where it all comes together: the hy-image-v3.5-preview API speaks the Chat/Messages protocol — the same paradigm as your LLMs. Concretely, your existing orchestration stack can drive an image model without any re-architecture.

The mechanics are crystal clear. You send a prompt, you get an image back, and to modify the result you feed the previous image back into the conversation history along with an instruction. The revision loop — re-injecting the image + history — is native, not a bolted-on hack. This is API design built for iterative editing, not one-shot generation.

The technical parameters deserve a closer look, per the official Tencent Cloud documentation (updated September 21):

• Resolution: automatic size determination, or an explicit size up to 4096×4096 (4K). Width and height between 256 and 8192 pixels, with a maximum area of 16,777,216 pixels — exactly a 4K square.
• Optional external search (use_search_tool): the model can lean on search to ground the visual in up-to-date information — relevant for product posters or branding.
• Rich inputs: up to 20 reference images, 20 MB maximum each.
• 12-hour temporary URLs: generated images expire, so download your assets immediately.

One small contradiction worth flagging, because it matters: TechFlow mentions output "up to 2K," while the API documentation allows 4K with an explicit size. The usual rule applies here: the technical documentation — dated and precise — takes precedence over the announcement post. But test it before promising 4K to your client.

The pricing, for its part, leaves no room for ambiguity: 0.15 RMB per 2K image — roughly $0.02 at September 2026 exchange rates — billed only for images actually generated. The API is also available via Tencent Cloud TokenHub, MPS, and VOD. At that price, image cost becomes a negligible line item in almost any production pipeline — which is exactly the point.

Qwen-Image-2.1: open-source as the storefront, the license as the moat

Alibaba played both sides during its conference: on one hand, spectacular closed announcements, on the other, an open-source release — Qwen-Image-2.1, a compact 7B-parameter model that claims, according to The Decoder, to outperform closed models in image generation.

The shipped package is smarter than a simple weight dump. Alibaba publishes two prompt-rewriting checkpoints, PE-T2I for text-to-image and PE-I2I for editing, fine-tuned from Qwen3.5-VL 9B. Their role: take your short prompts and expand them into rich prompts, which mechanically changes an image model's output quality. On the editing side, we're talking about up to 10 reference images, local mask-based edits, and support for transparent RGBA, according to Mixed News.

And that's where the shoe pinches. The historical Qwen-Image models were under Apache 2.0 — you could embed them in a commercial product without asking for permission, as the official GitHub repository reminds us. Qwen-Image-2.1 moves to the Qwen Research License Agreement: "FOR NON-COMMERCIAL PURPOSES ONLY". Research and evaluation, period.

My reading is straightforward: this is strategic open-washing. The open weights serve to attract researchers, generate community benchmarks, and feed the ecosystem — but the license precisely prevents the very audience that downloads open-source weights, developers, from turning them into a product. The moat, for its part, remains closed.

The nuance that partially saves the situation: the older Qwen-Image checkpoints remain under Apache 2.0. If you need open weights for commercial use, the door isn't closed — it has just moved to the previous versions. And if the open ecosystem interests you beyond licenses, note that Stanford just released GPIC, 28 trillion pixels for training image generation models: open data is moving forward too.

The dashboard: where are the Chinese models really?

Short answer: not in the top 10 of public image generation rankings — and that's precisely where all the ambiguity of this war resides. On one side, flattering internal claims; on the other, independent leaderboards where no Chinese model appears.

Here's the state of play of the public rankings as consulted to date (September 2026 — check the up-to-date rankings):

Three readings stand out. First, OpenAI dominates with a lead of nearly 130 points over second place — that's not a gap, it's a chasm. Next, Google saturates the ranking with three models, and the most interesting one isn't a Pro: it's nano-banana-2 with web search (1268), which outperforms its own Pro variants. Web grounding pays off, literally.

Finally, the Chinese absence. Tencent claims parity with Seedream 5.0 Pro and the edge over Nano Banana Pro — a model that sits here at 1242. Either the Chinese models aren't submitted to these rankings, or the independent evaluations haven't yet included them, or the internal claims are optimistic. All three hypotheses are defensible; none is verifiable today.

That's the central point of caution in this news story: China is fighting on internal benchmarks, the West on public arenas. Our comparison of the best image generation AIs will be updated as soon as an independent evaluation of Hy Image 3.5 is released — until then, any Chinese ranking is a statement, not a measurement. For broader context, AI Weekly follows these launches as they happen.

Three concrete scenarios to take advantage of them right now

These two launches aren't just for show: there are practical decisions to make this week, depending on your profile.

1. You manage e-commerce or marketing visuals. This is the declared hunting ground of Hy Image 3.5: posters, brand visuals, catalogs. With 5 reference images per request and 0.15 RMB per 2K image, generating 1,000 product variations costs about 150 RMB — less than 20 euros. At that price, the real cost is no longer generation, it's curation. To structure this workflow, see our AI tools for marketing.

2. You're a product or UI designer. Multi-turn conversational editing changes the way you work: you iterate on a mockup as if with a tireless junior colleague, feeding the previous image back in with a correction. Synchronous output makes this flow usable in an interactive tool — not in a batch pipeline. Storyboards and UI design are explicitly cited by Tencent as target use cases.

3. You're a developer and want to build on top of it. The Messages protocol does almost all the work: your existing LLM orchestration layer drives the image model. Two non-negotiable precautions: download the images immediately (URLs valid for 12 hours) and abstract the provider, since you're on a preview. To explore on a zero budget, our selection of free AI image tools does the job, and to publish a portfolio of your tests, simple hosting like Hostinger is more than enough.

❌ Common Mistakes

Mistake 1: Taking the "+30%" at face value

This figure comes from blind GSB tests run by Tencent's internal designers, based on comparisons against the previous version and competitors. No independent evaluation confirms it. The solution: build your own set of 20 to 50 prompts representative of your use cases, have your team evaluate the results blind, and let your own data speak.

Mistake 2: Building a product on a "Preview" API without a plan B

The model ID contains "preview" for a reason: behavior, pricing, and even availability can change without notice. Solution: isolate the provider behind an abstraction layer, keep your conversation histories portable (the Messages protocol helps), and keep a second model tested in parallel.

Mistake 3: Deploying Qwen-Image-2.1 in production without reading the license

The Qwen Research Agreement license is explicit: non-commercial use only. Integrating it into a paid product exposes you legally, even when self-hosting. Solution: if you need open weights for commercial use, fall back to the earlier Qwen-Image models under Apache 2.0, or accept a closed API.

Mistake 4: Ignoring temporary URLs and size limits

Generated images are served via URLs valid for 12 hours: if your pipeline doesn't download them immediately, you lose assets. And on the dimensions side, the maximum area of 16,777,216 pixels with sides between 256 and 8192 px requires upstream validation — otherwise, API errors in production.

❓ Frequently Asked Questions

Is Hy Image 3.5 Preview accessible from Europe?

The model runs through the Tencent Cloud API, whose documentation is public, and it is integrated into Tencent products such as Yuanbao and ima. Actual access from abroad depends on Tencent Cloud's commercial terms and the regional availability of its endpoints. Check availability in your region before starting any development.

How much does an image with Hy Image 3.5 actually cost?

The API is billed at 0.15 RMB per 2K image, roughly $0.02 at the September 2026 exchange rate, and only for images actually generated. Usage-based billing, with no announced minimum, makes the marginal cost nearly zero for testing. Check the current pricing on Tencent Cloud before committing to any volume.

Can Qwen-Image-2.1 be used in a commercial product?

No. The model is released under the Qwen Research License Agreement, explicitly limited to non-commercial use — research and evaluation only. This is a break from the Apache 2.0 licensing of previous Qwen-Image models, which remain usable commercially. For a product, your open-weight option is therefore the earlier checkpoints.

Does Hy Image 3.5 Preview beat Nano Banana Pro?

According to Tencent's internal GSB tests, slightly — but no independent evaluation confirms this at this stage. On the public leaderboards consulted in September 2026, nano-banana-pro scores around 1242, with no Chinese model in the top 10. The only reliable answer: test both on your own use cases.

What is the GSB methodology mentioned by Tencent?

GSB stands for Good/Same/Bad: evaluators compare two images produced from the same prompt, without knowing who generated them, and vote "better", "equal", or "worse". It's a solid methodology for measuring progress between two versions of the same model, but far less reliable for comparing two competing vendors.

✅ Conclusion

In a single day, September 22, 2026, China showcased the three weapons of its generative-image war: timing calibrated to the competition, slashed prices, and a tightening license — now the ball is in your court: test things for yourself rather than trusting internal benchmarks; to choose the model best suited to your needs, start with our comparison of the best AI image generators and follow the latest AI news.