Hy4 preview vs GLM-5.3-Flash: Tencent's Text-Only Flagship or Zhipu's Multimodal Flash?
Compare Tencent's Hy4 preview and Zhipu's GLM-5.3-Flash for production LLM workloads — coding agents, cost, context, and multimodal input.
Short Answer
Compares Tencent's Hy4 preview (text-only flagship) and Zhipu's GLM-5.3-Flash (native-multimodal Flash tier), drawing on launch reporting, official docs, hands-on reviews, and public pricing. VERSIONS COMPARED: Hy4 preview (770B-A49B, text-only) · GLM-5.3-Flash (320B-A18B, native multimodal)
Our Recommendation
Blind-test winner over GLM-5.3 (2.99 vs 2.92), DeepSWE jumped 28→64.3, 1M context, Apache 2.0 — the stronger pure-text engine.
Only one of the two accepts images/video/files natively, and it costs ~1/7 as much per token — the clear production value pick.
Hy4 preview is text-only; any vision/file-input task must go to GLM-5.3-Flash or another multimodal model.
NOT A STRICT EITHER/OR — THE TWO CAN BE COMBINED
Comparison at a Glance
| Dimension | Hy4 preview | GLM-5.3-Flash |
|---|---|---|
| Total / active params | 770B / 49B (MoE) | 320B / 18B (MoE) |
| Context window | 1M tokens (1024K) | 1M tokens (128K max output) |
| Input modalities | Text only | Native: image, video, file, text |
| API cost (input) | ~$0.83/M (≈¥6/M) | ~¥0.8/M (≈$0.11/M, ~1/10 of GLM-5.3) |
| API cost (output) | ~$2.50/M (≈¥18/M) | ~¥2.8/M (≈$0.39/M) |
| License | Apache 2.0 | MIT (reported) |
| Activation ratio | ~6.4% (49B active) | ~5.6% (18B active) |
| Coding / agent anchor | Blind-test 2.99 vs GLM-5.3 2.92 · DeepSWE 28→64.3 | AA Index 57, on par with Claude Opus 4.8 |
| Ecosystem | CodeBuddy / WorkBuddy / Yuanbao / ima / TokenHub | OpenRouter leader · runs on domestic chips |
SOURCE: Tencent Hunyuan release & hands-on (AUG 2026) · Zhipu open docs · IT之家/Sohu/Sina · lmmarketcap pricing. Hy4 preview is a flagship-tier model; GLM-5.3-Flash is the Flash-tier sibling of GLM-5.3.
Why Each One Wins
Coding & Agentic Workflow
- Hy4 preview: Tencent's internal 163-expert blind test over 203 engineering tasks scored it 2.99/4 vs GLM-5.3's 2.92 and Kimi K3's 2.94; DeepSWE jumped from 28 to 64.3, and it is tuned with CodeBuddy/WorkBuddy.
- GLM-5.3-Flash: Scores 57 on the Artificial Analysis Intelligence Index (on par with Claude Opus 4.8) and tops Z.ai's own Code Bench — remarkably strong for a Flash-tier model.
Multimodal Input
- Hy4 preview: Text-only. Hands-on testing confirms no vision or multimodal input — any image/video/file task needs a different model.
- GLM-5.3-Flash: Native multimodal: accepts image, video, file, and text input — the first GLM-5 family model with this baked in.
Context Window
- Hy4 preview: 1024K (1M-token) context, up from Hy3's 192K — built for very long documents and long-running agent sessions.
- GLM-5.3-Flash: Also 1M-token context with 128K max output — matches Hy4 on context length.
Cost & Efficiency
- Hy4 preview: ~$0.83/M input and ~$2.50/M output on TokenHub/lmmarketcap — flagship-tier pricing, roughly 7× GLM-5.3-Flash per token.
- GLM-5.3-Flash: ~¥0.8/M input and ~¥2.8/M output (~1/10 of GLM-5.3, limited-time ~1/20), with a more aggressive 5.6% activation ratio for lower serving cost.
License & Ecosystem
- Hy4 preview: Apache 2.0 with explicit patent grant — enterprise-friendly; deeply integrated into CodeBuddy, WorkBuddy, Yuanbao, and ima, and served via Tencent Cloud TokenHub.
- GLM-5.3-Flash: MIT reported (open weights), first launched as 'Ox Alpha' and topped OpenRouter call volume on day one; runs on domestic chips.
Trade-offs
Choose Hy4 preview, you give up…
- No image/video/file input at all — text-only for now
- Roughly 7× the token price of GLM-5.3-Flash
- Preview-stage model: production stability and API behavior are still maturing
Choose GLM-5.3-Flash, you give up…
- Blind-test ceiling below Hy4 preview (2.92 vs 2.99 vs GLM-5.3)
- Smaller active-parameter headroom (18B vs 49B) for the hardest reasoning
- Weaker first-party integration if you live in Tencent's CodeBuddy/WorkBuddy ecosystem
Supporting Evidence
2026-08
2026-08
2026-08
2026-08
2026-08
2026-08
2026-08
2026-08
Limitations & Known Caveats
- Both models launched within days of each other (Hy4 08-28, GLM-5.3-Flash 08-26); benchmarks, pricing, and production behavior are very new and may shift.
- Hy4 preview is a preview-stage, text-only model — no multimodal input and no long production track record yet.
- The 2.99 vs 2.92 blind test is Tencent's own internal evaluation (163 experts, 203 engineering tasks); treat it as vendor-reported.
- GLM-5.3-Flash pricing (~1/10 of GLM-5.3) is vendor-reported with a limited-time discount; exact per-token rates vary by plan and region.
- Hy4 preview prices are USD quotes from lmmarketcap; CNY estimates use ~7.2 exchange rate and are approximate.
- AA Index 57 (GLM-5.3-Flash) comes from vendor and press claims, not an independent universal benchmark.
Alternatives
GLM-5.3 · Qwen3.8-Flash · Kimi K3 · DeepSeek-V4-Flash · Claude Opus 4.8
Bottom Line
Which should you choose?
- You build text-first coding agents or long-running autonomous workflows and want peak intelligence.
- You need a 1M-token context and value Apache 2.0's explicit patent grant.
- You live in Tencent's CodeBuddy/WorkBuddy ecosystem and want the model it was tuned with.
- Your workload never needs image/video/file input.
- Your application must accept images, video, or files as input.
- Cost-per-token is a dominant constraint and you run high-volume calls.
- You want frontier-tier quality (AA 57) at Flash-tier serving economics.
- You prefer the open-source/OpenRouter community momentum and domestic-chip deployment.
Frequently Asked Questions
Can Hy4 preview process images or files?
No. Hy4 preview is text-only; for image/video/file input you must use GLM-5.3-Flash or another multimodal model.
Which is cheaper to call via API?
GLM-5.3-Flash, by roughly 7× — about ¥0.8/M input vs Hy4 preview's ~$0.83/M (≈¥6/M).
Which has the longer context window?
Both support 1M tokens. GLM-5.3-Flash adds a 128K max output budget.
Which is better for AI coding agents?
For pure text/coding power, Hy4 preview wins Tencent's blind test (2.99 vs GLM-5.3's 2.92). GLM-5.3-Flash is still exceptional value and can also watch UI/video input in the loop.
Is either model open source?
Both. Hy4 preview is Apache 2.0; GLM-5.3-Flash is MIT (reported), having topped OpenRouter call volume on its first day.
How Airdix Makes Decisions
Independence guarantee: recommendations are based only on criteria and evidence. A product can pay for visibility, but never for a recommendation. Every claim is traced to a source and marked with a verification date. Recommendation ≠ Paid Placement
Compares Tencent's Hy4 preview (text-only flagship) and Zhipu's GLM-5.3-Flash (native-multimodal Flash tier), drawing on launch reporting, official docs, hands-on reviews, and public pricing.
Params, context, pricing, and benchmarks from Tencent Hunyuan release coverage, CSDN hands-on, Zhipu open docs, IT之家/Sohu/Sina coverage, and lmmarketcap (Aug 2026).
Hy4 preview released 2026-08-28, GLM-5.3-Flash released 2026-08-26; no independent side-by-side reproduction performed here.
Tencent 163-expert blind test · Artificial Analysis Intelligence Index · Z.ai Code Bench · DeepSWE
Run the same agentic-coding and multimodal tasks on both APIs and compare output quality, latency, and your token bill.
Very new launches; vendor-sourced blind-test and pricing claims; Hy4 preview pricing is USD-based and approximate in CNY.
AUG 2026 (content quality review)