Tencent Hy4 vs GLM-5.3: Which Chinese AI Model Is Better for Coding?

Two weeks apart, China's two big flagships landed: Zhipu's GLM-5.3 (August 14) with the year's biggest published long-horizon coding jumps, and Tencent's Hy4 preview (August 28) with the largest open-weights spec yet — and the two vendors point at opposite kinds of evidence. GLM-5.3 publishes full benchmark tables; Hy4 preview publishes a blind expert evaluation and no public numeric scores. This comparison works through what each side actually shows for coding, long-context, reasoning, speed, cost, openness, and developer usability — without inventing test scores neither vendor released.

Quick Comparison

Contenders at a Glance

GLM-5.3 is Zhipu's coding & cybersecurity flagship — an unusual release built on the same base as GLM-5.2 with all gains from scaled post-training, posting the largest long-horizon coding jumps in the GLM line's history (Terminal-Bench 3.0: 4.6 → 28.3) plus an emergent cyber capability. Its open weights are held back for safety evaluation. Full details in our GLM-5.3 review.

Hy4 preview is Tencent's new-generation MoE flagship: 770B total / 49B active, a 1M+ context, Apache-2.0 weights released at launch, and — per Tencent's announcement — the first model where the company described a "recursive self-improvement loop" (the model participated in auto-optimizing its own training data strategy and evaluation). It arrived as a preview — Tencent promises more Hy4 models. Full details in our Hy4 preview review.

Coding: Benchmarks vs. Blind Eval GLM by published data, Hy4 by expert preference

This is the crux of the comparison, and the two sides cannot be directly reconciled:

• Zhipu published tables (vendor-reported): Terminal-Bench 2.1 88.2, Terminal-Bench 3.0 28.3, DeepSWE v1.1 66.9, Agents' Last Exam 28.5, and 31.4% completion on the internal Z.ai Code Bench at ~50K output tokens (vs Claude Opus 4.8's 29.5% at 120K and Claude Fable 5's 39.5% at Max).
• Tencent published an expert evaluation (vendor-run): 163 experts, 203 engineering tasks, blind scoring — Hy4 preview 2.99/4 vs GLM-5.3's 2.92/4 (46.8% wins, 12.8% ties, 40.4% losses) and vs Kimi K3's 2.94/4 (51.2% wins).

The honest synthesis: if you want numbers, GLM-5.3 is the only one with them; if you want engineering-expert preference, Tencent's own blind test puts Hy4 ahead of GLM-5.3 by a small margin. Both are vendor-generated, neither is independently verified, and the two evidence types measure different things (task completion rates vs. expert preference). Notably, Hy4 preview's expert edge over GLM-5.3 on engineering tasks, if it holds up in independent testing, would make Tencent's preview the more interesting coding pick despite the missing tables.

Long-Context Tie (1M-class both)

Both models are 1M-token class: GLM-5.3's evaluation footnotes reference 1M-token runs (NL2Repo) and its GLM-5.2 lineage carried a million-token window; Tencent's newsroom states Hy4 preview's context as "over 1M." GLM-5.3's post-training specifically targeted long-horizon coding agents with days-scale engineering work units, while Hy4's compressed Gated DSA attention (IndexCache) is designed to make 1M context tractable at 770B scale. Neither vendor publishes clean per-request latency at 1M, so treat long-context reliability as unverified on both sides. For the wider long-context field, see Best AI Models for Long-Context Work 2026.

Reasoning Unverified both

GLM-5.3's published reasoning-adjacent results: HLE with tools 62.5, GDPval-AA v2 1,769 (~4% behind Claude Opus 5 Max per Zhipu's table). Hy4 preview's reasoning evidence is the blind expert eval plus Tencent's positioning (recursive self-improvement pipeline, strong RL). No public numeric reasoning benchmark exists for Hy4 preview as of August 30, 2026, so a direct reasoning comparison is not possible — we won't invent one.

Speed Unverified both

Neither vendor published standard inference latency for these models. What exists: Zhipu reports system-level optimizations that improved long-horizon RL training throughput by 2.3x+ (training, not inference), and Tencent reports a 31.8% end-to-end inference throughput gain from auto-optimized kernels and operator fusion versus its own baseline — plus native MTP speculative decoding in Hy4. Meanwhile Tencent's repo warns Hy4 preview can spend longer than necessary reasoning and over-verify its own work, which directly raises latency and token cost on simple coding tasks. GLM-5.3's always-on thinking has a similar effect (no disabled-thinking path). For speed-sensitive coding agents, both need load testing; neither ships published latency data.

Cost Hy4 published, GLM unknown

Hy4 preview has real published pricing: ¥6 per million input tokens, ¥18 per million output tokens, cache hits from ¥0.3 (≈$0.85 / $2.54 / $0.04 at ~7.1 CNY/USD — our conversion, marked as estimate). GLM-5.3 has no published standalone API price; it runs through the GLM Coding Plan's points-based quota (50% off-peak consumption) and ZCode with a 1.5x quota boost through August 31. A direct per-token comparison is impossible today — if you need line-item pricing, Hy4's is public; GLM's is a subscription calculus. For context, Zhipu's Flash-tier sibling GLM-5.3-Flash costs $0.15/$0.50 — an order of magnitude below Hy4's published rates.

Openness & Deployment Hy4 preview — weights available today

This dimension has a clear winner today: Hy4 preview's Apache-2.0 weights were released at launch on Hugging Face, ModelScope, GitCode, and CNB, with official vLLM and SGLang images, a finetuning pipeline, and AngelSlim quantization tooling — fully self-hostable right now. GLM-5.3's weights were not yet confirmed released at the time of this audit: Z.ai's original timeline pointed to approximately two weeks after the August 14 launch, pending security hardening (license unannounced) — verify current status on z.ai. If open weights are your requirement today, Hy4 preview is the one you can run now — the same conclusion our Best Open-Weight AI Models for Coding 2026 draws.

Developer Usability Depends on your stack

• GLM-5.3: day-one rollout to all GLM Coding Plan subscribers; first-party ZCode harness; API model ID glm-5.3 with always-on thinking and reasoning_effort low/high/max; documented breaking change for pipelines that disabled thinking.
• Hy4 preview: first release across WorkBuddy/CodeBuddy (domestic + international), Yuanbao, and ima; APIs on Tencent Cloud Tokenhub and OpenRouter; OpenAI-compatible endpoint after self-hosting; recommended sampling temperature 0.9 / top_p 1.0 with reasoning_effort: "no_think" for direct replies.

If you live in coding agents today, GLM-5.3 is the more turnkey path (Coding Plan + ZCode). If you self-host or use OpenRouter-style routing, Hy4 preview's open weights and OpenAI-compatible serving are the practical choice. For agent harness context, see Best AI Coding Assistant 2026.

Which One Should You Choose?

• Choose GLM-5.3 if: you want published benchmark numbers, you're on the GLM Coding Plan, or you're tracking the open-weights release for when it lands.
• Choose Hy4 preview if: open Apache-2.0 weights today matter more than benchmark tables, you trust expert-preference evidence, or you're in Tencent's product ecosystem (WorkBuddy/CodeBuddy free trial).
• Choose neither for peak coding if: closed frontier models are on the table — per Zhipu's own table, GPT-5.6 Sol and Claude Fable 5 lead the hardest suites (see GPT-5.6 vs Claude).
• Wait and re-test if: Hy4's benchmark tables land (the full Hy4 line is promised) or GLM-5.3's weights drop — both events will change this comparison materially.

Final Verdict

FAQ

Which is better for coding: Hy4 preview or GLM-5.3?

GLM-5.3 has the only published coding benchmarks (TB2.1 88.2, DeepSWE 66.9, vendor-reported); Hy4 preview has a vendor-run blind eval edging GLM-5.3 (2.99 vs 2.92) but no public numeric scores. No independent comparison exists.

Does Hy4 preview beat GLM-5.3?

On Tencent's blind expert eval, yes (2.99 vs 2.92). On published benchmarks, GLM-5.3 is the only one with any. Neither is independently verified.

Which has a bigger context window?

Both 1M-class: Hy4 preview "over 1M" per Tencent's newsroom; GLM-5.3 1M per evaluation footnotes. No clean spec number published for GLM-5.3.

Which is more open?

Hy4 preview — Apache-2.0 weights released at launch. GLM-5.3's weights were not yet confirmed released at the time of this audit (Z.ai timeline: ~2 weeks after the Aug 14 launch) — verify current status on z.ai.

Which is cheaper?

Hy4 preview has published pricing (¥6/¥18 per million); GLM-5.3 has no standalone API price — only the Coding Plan points system. Not directly comparable.

Sources (accessed August 30, 2026): Zhipu official GLM-5.3 announcement and published evaluations (as documented in our GLM-5.3 review), Tencent official newsroom release, Tencent-Hunyuan/Hy4-preview GitHub repository. All benchmarks vendor-reported and reproduced as published; CNY→USD conversions approximate at ~7.1 CNY/USD. ToolStep claims no laboratory test results. See our editorial policy.