GMI Cloud Says Tencent Hy3 Beat GLM And DeepSeek In A Small Web Design Test
The GPU cloud provider tested four open models on three webpage prompts, a narrow but useful read on where design coding is moving.
By Ryan Merket · Published
Why it matters
GMI Cloud's small test shows open models are being judged by task-level usefulness - especially frontend output - instead of broad benchmark rank alone.
GMI Cloud (@gmi_cloud) said on July 7th that Tencent's Hy3 produced stronger visual webpage output than GLM 5.1 and DeepSeek V4 in a small internal design test, putting another Chinese open model into the fight for developer workflows that used to default to closed frontier systems.
The test was narrow by design. GMI Cloud said it ran Tencent Hy3, GLM 5.2, GLM 5.1 and DeepSeek V4 through three webpage design prompts at "medium effort," then compared the visual output. In the post on X, GMI Cloud wrote that Hy3 "outperformed both GLM 5.1 and DeepSeek V4" on visual output and described the model as "great with design" and as affordable as DeepSeek V4.
https://x.com/gmi_cloud/status/2074303573223993636
That is a vendor-run comparison, not an independent benchmark. GMI Cloud did not publish a full scoring rubric in the post, did not identify the three prompts in text, and did not disclose whether the outputs were judged by humans, an automated evaluator, or a fixed checklist. The result is still worth watching because it points at a practical shift in model selection: developers are starting to route frontend-generation tasks model by model, rather than treating open models as a single commodity tier behind Claude, Gemini or GPT.
GMI Cloud has an obvious commercial reason to publish this kind of comparison. The company sells AI infrastructure and inference access, and its site describes GMI Cloud as an "AI-native infrastructure platform" spanning serverless APIs, orchestration, dedicated GPU compute and NVIDIA H100, H200 and Blackwell hardware. Its pricing page lists H100 GPUs from $2.00 per GPU-hour, H200 from $2.60 per GPU-hour, B200 from $4.00 per GPU-hour and GB200 from $8.00 per GPU-hour. A public test that names cheap, capable open models helps GMI Cloud make the case that customers can build production AI systems without locking every high-volume task to expensive closed APIs.
The more interesting piece is Hy3. Tencent introduced Hy3 preview this spring as an open-sourced model available through GitHub, Hugging Face, ModelScope and GitCode, with support for inference frameworks including vLLM and SGLang. Tencent said at launch that Hy3 preview was tuned for agent workflows, coding, search execution and long-context instruction following, and that it could run through agent frameworks including OpenClaw, OpenCode and KiloCode.
Tencent also attached pricing and efficiency claims to the release. The company said Hy3 preview delivered a 40% improvement in inference efficiency, and listed TokenHub pricing starting around $0.18 per million input tokens, $0.06 per million cached input tokens and $0.59 per million output tokens. Those are Tencent-supplied numbers, but they explain why a provider such as GMI Cloud would test the model against DeepSeek V4 and GLM: the contest is becoming as much about cost per acceptable UI as leaderboard placement.
Hy3's public model card on Hugging Face shows the model is available as tencent/Hy3-preview, tagged for text generation, Transformers and safetensors. The same page includes example deployment paths for Transformers, vLLM, SGLang and Docker Model Runner. That matters for startups because a visually competent webpage-generation model that can be self-hosted or routed through multiple providers gives engineering teams more control over latency, cost and data handling than a single closed API.
DeepSeek V4 remains the benchmark GMI Cloud appears to be measuring against. In April, GMI Cloud published its own DeepSeek V4 test and said the model launched with a 1 million token context window, V4-Pro and V4-Flash variants, OpenAI-compatible access through GMI Cloud, and model IDs that could be used inside tools such as Cursor. GMI Cloud also wrote that V4-Pro cost $1.74 per million input tokens and $3.48 per million output tokens, while V4-Flash cost $0.14 per million input tokens and $0.28 per million output tokens. Those figures came from GMI Cloud's own blog, so they should be treated as provider-reported pricing rather than a permanent market rate.
GLM 5.2 gives the comparison another axis. Z.ai describes GLM 5.2 as its latest flagship model for long-horizon tasks with a 1 million-token context window and stronger performance over GLM 5.1. Z.ai's GitHub repository highlights 1 million-token context, stronger coding capability and multiple thinking-effort levels. OpenRouter lists GLM 5.2 at $0.9086 per million input tokens and $2.856 per million output tokens, with a 1 million-token context window and a June 16th, 2026 release date on its model page.
That leaves GMI Cloud's July 7th post as a useful datapoint rather than a verdict. Three prompts can expose obvious frontend weaknesses - layout drift, poor spacing, broken component hierarchy, weak styling taste - but they cannot settle model quality across design systems, accessibility, interaction logic, browser compatibility or production code maintainability. GMI Cloud's phrasing is careful on one point: it says Hy3 outperformed on visual output, not that Hy3 is the best overall coding model.
For founders building developer tools, design agents or internal app generators, that distinction is the product decision. A model that produces a better first visual pass can cut iteration time even if another model is stronger at repo-scale reasoning or multi-file refactors. The likely winning stack routes tasks: one model for UI draft generation, another for long-context codebase work, another for cheap high-volume edits. GMI Cloud's post is a small public example of that routing logic arriving in ordinary product development.
The open-model market is also compressing fast. Tencent, DeepSeek and Z.ai are each pushing open or open-weight systems into agentic coding, long context and low-cost inference. Their claims differ, their licenses and deployment paths are not identical, and vendor benchmarks remain uneven. The common pressure is clear: if open models can handle enough frontend generation at a fraction of closed-model cost, infrastructure providers gain usage, startups gain margin, and closed labs lose some pricing power at the application layer.