Hy4 Preview vs GLM 5.3 vs Kimi K3: What the Early Evidence Shows

01

The short version

Hy4 Preview stands out for Apache-2.0 weights, official full/FP8 checkpoints, a 1M-token context and documented vLLM/SGLang deployment. GLM 5.3 is integrated with Z.ai's coding ecosystem, while Kimi K3 is positioned for long-horizon coding, knowledge work and reasoning.

There is no defensible universal winner. The best choice depends on your workload, deployment model, serving budget and need for infrastructure control.

02

What Tencent's internal comparison says

Tencent says 163 internal experts evaluated outputs across 203 engineering tasks. It reports average scores of 2.99 for Hy4, 2.92 for GLM 5.3 and 2.94 for Kimi K3.

The reported head-to-head results are 46.8% wins / 12.8% ties / 40.4% losses against GLM 5.3, and 51.2% wins / 7.9% ties / 40.9% losses against Kimi K3.

03

Context, access and self-hosting

Tencent and Moonshot both document 1M-token context for Hy4 and Kimi K3. Long context is useful for large repositories and file collections, but serving cost, KV-cache demand and retrieval quality still determine whether the maximum is practical.

Hy4 has official Apache-2.0 full and FP8 checkpoints. Kimi currently provides first-party hosted API access and lists K3 pricing at $3 per million input tokens, $15 per million output tokens and $0.30 per million cache-hit tokens. Z.ai's current first-party pricing lists GLM 5.3 at $1.40 input, $4.40 output and $0.26 cached input per million tokens.

04

Which model should you use?

Choose Hy4 when official open weights, Apache licensing, self-hosting and model-level infrastructure control are central. See the local deployment guide before estimating hardware.

Consider GLM 5.3 when Z.ai's coding ecosystem and current first-party API economics fit the workflow. Consider Kimi K3 where its native multimodal, 1M-context agent stack and hosted API are a stronger operational match.

For production, benchmark the candidates on the same tasks, prompts, tools, context size and evaluation rubric. Vendor comparisons cannot replace workload-specific testing.