2026年8月19日星期三

Zhipu GLM-5.3 Review 2026: Post-Training Only, Open-Source Coding #1, AA 60, CyberGym #1, $1.40/M

Zhipu GLM-5.3 Review 2026: Post-Training Only, Open-Source Coding #1, AA 60, CyberGym #1, $1.40/M

Zhipu GLM-5.3 Review 2026: Post-Training Only, Open-Source Coding #1, AA 60, CyberGym #1, $1.40/M

Verdict first: GLM-5.3 is the strongest argument yet that post-training scaling beats scaling parameters. Zhipu kept the exact same 743B MoE base as GLM-5.2 and still lifted coding by ~50% — Terminal-Bench 3.0 jumps from 4.6 to 28.3 (open-source #1), DeepSWE hits 66.9 (past DeepSeek V4 Pro), and an unexpected cybersecurity ability emerges with CyberGym at 84.5% (global #1). Its Artificial Analysis Intelligence Index of 60 ties Kimi K3 for the top open-source score, at the lowest single-task cost among frontier flagships. If you code or run agents and want open weights near the frontier, this is the one to watch.

Zhipu GLM-5.3 review 2026 — 743B MoE, post-training only, open-source coding #1, CyberGym #1

GLM-5.3 — Zhipu's open-source coding flagship, powered purely by post-training

What Is GLM-5.3?

Released on August 14, 2026 with its API going live August 19, GLM-5.3 is Zhipu AI's (Z.ai) latest flagship — and its most provocative. The architecture and parameter count are identical to GLM-5.2 (743B MoE, ~40B active). The entire leap comes from post-training scaling, a move Zhipu summarizes bluntly: "Scaling post-training is all we did."

The result: open-source #1 coding, a frontier-band Intelligence Index of 60, and an unexpected emergence of cybersecurity capability that has security teams paying attention.

Specifications at a Glance

ItemValue
VendorZhipu AI (Z.ai)
Release / API liveAugust 14 / August 19, 2026
PositioningOpen-source coding flagship, post-training only
ArchitectureSparse Mixture-of-Experts
Total / active params743B / ~40B active
Context window1M input / 128K output
InputsText only
API price$1.40 / $4.40 per M tokens (unchanged from 5.2)
Open weights~Aug 28, MIT license
Throughput~145–168 tokens/sec

Pricing: Cheapest Frontier Cost

ItemGLM-5.3Note
Input$1.40 / M tokensUnchanged from GLM-5.2
Output$4.40 / M tokens
Single-task costLowest among frontier flagshipsArtificial Analysis

While DeepSeek and others are raising prices, GLM-5.3 holds its rate. And its token efficiency is the kicker: in High mode, GLM-5.3 scored 31.4% using ~50K tokens, while Claude Opus 4.8 needed ~120K tokens for 29.5% — half the tokens, higher completion. For heavy users, that compounds into dramatically lower bills.

Benchmarks: Post-Training Explosion

Benchmark5.25.3Highlight
Terminal-Bench 3.04.628.36x, open-source #1
DeepSWE v1.146.266.9Beats DeepSeek V4 Pro (62.7)
Agents' Last Exam (CLI)23.828.5Beats Kimi K3 & Opus 4.8
CyberGym (vuln discovery)77.284.5Global #1 (>GPT-5.6 Sol 83.6)
ExploitBench24.454.42x+
SWE-Marathon v1.119.442.5~2x
AutomationBench26.248.2
FrontierSWE67.578.1
GDPval-AA v2 (Elo)15081769Highest in table

Artificial Analysis Intelligence Index: 60 — the frontier band alongside Claude Fable 5 and GPT-5.6 Sol, and tied with Kimi K3 for the top open-source score.

Why "Post-Training Only" Matters

GLM-5.3 is the cleanest proof that training technique beats raw scale. No new architecture, no added parameters — just a massively scaled post-training pipeline built on an "AI researcher writes questions, AI grader verifies" synthetic loop, with long-horizon coding training sped up 2.3x. Chinese labs are clearly shifting from "scale at any cost" to "squeeze more from less."

And then there's the security surprise: before launch, a red-team review with Tsinghua and Nankai found 2,436 vulnerabilities in 269 open-source projects (1,097 medium-to-high severity) — including a 40-year-old DNS protocol-level flaw with 10M+ public DNS services potentially exposed. That capability is exactly why the weights are delayed to Aug 28 for a security review.

Pros & Cons

Pros

  • ✅ AA 60 — tied for the top open-source Intelligence Index
  • ✅ Open-source #1 coding; ~50% better than 5.2
  • ✅ CyberGym global #1 — unexpected security capability
  • ✅ Outstanding token efficiency, lowest frontier single-task cost
  • ✅ 1M context + MIT open weights (Aug 28)
  • ✅ Price unchanged in a market of hikes

Cons

  • ❌ Weights not out until Aug 28 — can't self-host yet
  • ❌ Text-only, no multimodal
  • ❌ Stock dropped on launch day (lukewarm market reaction)
  • ❌ Some figures are Zhipu-internal until weights are verified
  • ❌ Hardest reasoning tasks still trail closed frontier models

Who Should Use It (And Who Shouldn't)

Great fit

  • Developers: long-horizon coding, terminal automation, codebase-level tasks
  • Agent builders: multi-step tool calls, real software engineering
  • Security researchers: vulnerability discovery and red-team work
  • Cost-conscious teams wanting near-frontier open models

Poor fit

  • Multimodal applications (text-only)
  • Strict private deployment before Aug 28 weights
  • Creative / long-form writing (not its strength)

Frequently Asked Questions

Is GLM-5.3 open source?

Weights release ~Aug 28 under MIT on Hugging Face (zai-org), delayed for security review. API is live now.

How big is it?

743B total MoE params, ~40B active, same base as GLM-5.2 — all gains from post-training.

How much does it cost?

$1.40/M input, $4.40/M output — unchanged from 5.2; lowest single-task cost among frontier flagships.

Is it really the best open coding model?

Terminal-Bench 28.3 is open-source #1; DeepSWE 66.9 beats DeepSeek V4 Pro; internal coding ~50% better than 5.2.

Is the cybersecurity ability real?

CyberGym 84.5% global #1; red-team found 2,436 vulns in 269 projects, incl. a 40-year DNS flaw affecting 10M+ services.

What is the context window?

1M input / 128K output, text-only.

How does it compare to Claude Fable 5 / GPT-5.6?

AA Index 60 puts it in the frontier band; it ties Kimi K3 for top open score and is far more token-efficient.

Sources

没有评论:

发表评论

Microsoft MAI Review 2026: The 7-Model Family That Says Goodbye to OpenAI — Trillion-Parameter Flagship, Now Default in GitHub Copilot

Microsoft MAI Review 2026: The 7-Model Family That Says Goodbye to OpenAI — Trillion-Parameter Flagship, Now Default in GitHub ...