GLM-5.3 Ties Kimi K3 on the AA Intelligence Index — Without a New Base Model

Update — August 28, 2026: Z.ai has released GLM-5.3 as open-weight. The weights are live on Hugging Face at zai-org/GLM-5.3, with a technical write-up at z.ai/blog/glm-5.3, and Z.ai calls it "our most capable model for agentic coding and cyber defense is now available to download, run, and customize." The release followed a roughly two-week safety review — about two weeks behind GLM-5.2's day-one weights drop. The Intelligence Index tie described below now holds between two downloadable models, though GLM-5.3's CyberGym and Terminal-Bench figures stay provider-reported until independent labs rerun them against the actual weights. Details and local-run options are in the "Can you run GLM-5.3 locally now?" section.

Update — August 28, 2026: Tencent's Hy4 preview (770B/49B active, 1M context, Apache 2.0) self-reports 2.99/4 against GLM-5.3's 2.92 on an internal blind eval — a margin inside the noise, but it prices below GLM-5.3 at $0.834/$2.501 per million tokens.

Kimi K3 has held the top Intelligence Index score among open-weights models since Moonshot AI released its weights on July 26, 2026. On August 18, 2026, Artificial Analysis published a fresh evaluation of GLM-5.3 that puts it at exactly the same number: 60.

That's the headline two duplicate wire posts ran with on August 19 — "GLM-5.3 Ties Kimi K3 as Most Intelligent Open Model With 60 Score" and a near-identical variant emphasizing the foundation model stayed unchanged. Both are accurate, but the second one is the more interesting story: Z.ai didn't retrain GLM-5.3 from scratch to close the gap with Kimi K3. It kept the same 753-billion-parameter base it shipped with GLM-5.2 and pushed the score up seven points through post-training alone.

TL;DR

What is the Artificial Analysis Intelligence Index, exactly?

The number everyone is citing — 60 — comes from Artificial Analysis's Intelligence Index, an independent composite score that averages a model's results across nine underlying evaluations: agentic real-world work tasks, agentic tool use, terminal coding, scientific reasoning and knowledge, graduate-level science questions, physics reasoning, knowledge reliability and hallucination resistance, and long-context reasoning. It is not a single leaderboard entry and not a benchmark Z.ai ran itself — Artificial Analysis evaluates every model independently, on its own infrastructure, using the same methodology across providers.

That independence is what makes the tie notable. Z.ai's own launch materials for GLM-5.3 emphasized CyberGym and defensive cybersecurity benchmarks; the Intelligence Index score is a separate, third-party read taken four days after launch, on August 18, 2026.

Claude Opus 5 still leads the overall field at 63 — three points ahead of the tied open-weights pair. That gap matters: GLM-5.3 closed distance to the frontier, it didn't reach it.

The benchmark comparison table

Sub-scores tell a more textured story than the tied headline number. On GDPval-AA v2 — Artificial Analysis's real-world agentic knowledge-work evaluation — GLM-5.3's Elo jumped from 1,524 (GLM-5.2) to 1,770, putting it second overall behind only Claude Opus 5 (1,855) and more than 100 points ahead of Kimi K3 (1,668). On AA-Omniscience, which measures factual reliability and resistance to hallucination, Kimi K3 still leads open-weights models at 20 against GLM-5.3's 14 — up from GLM-5.2's 4, but still trailing.

So "ties Kimi K3" is true at the aggregate level and false at the sub-score level in both directions. That's normal for composite indices, and it's exactly the kind of nuance a single headline score erases — see explainx.ai's complete guide to AI benchmarks for more on why aggregate scores flatten real per-task differences.

Zoom out one more level and the ranking tells a policy story, too. Three of the four models within a few points of Claude Opus 5 on this index — GLM-5.3, Kimi K3, and Qwen3.8 Max — come from Chinese labs, all released or updated within the same five-week window in July and August 2026. That clustering is the practical version of the distillation and export-control debate explainx.ai has tracked through the year: whatever the policy fight settles, the gap between frontier closed models and the best open-weights alternatives keeps shrinking on independent measurement, not just on vendor marketing claims.

What changed if the base model didn't?

This is the detail that separates a real engineering story from a routine leaderboard shuffle. Z.ai confirmed GLM-5.3 runs on the identical base model as GLM-5.2 — same 753B total parameters, same 40B active parameters, same 1M-token context window. There was no new pretraining run.

The seven-point Intelligence Index gain, and the much larger jumps on specific agentic benchmarks — Terminal-Bench 3.0 went from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9 — came entirely from post-training. Specifically, Z.ai scaled reinforcement learning on long-horizon task environments: more environments, more task diversity, more compute, all on the same RL training stack it had already built for GLM-5.2.

That's the practical takeaway for anyone evaluating open models: a lab can extract a full generation's worth of intelligence gain out of an existing checkpoint through RL scaling alone, without touching the architecture or running a new multi-month pretraining cycle. It's the same broader pattern explainx.ai covered when Kimi K3 first shipped as an open-weights frontier model with heavy post-training investment, and it echoes the RL-focused workflow Cognition described for SWE-1.7. Post-training is where a growing share of frontier competition is actually happening, not pretraining scale.

Can you run GLM-5.3 locally now?

Yes — as of August 28, 2026, Z.ai released GLM-5.3's open weights on Hugging Face at zai-org/GLM-5.3, with a technical write-up at z.ai/blog/glm-5.3. The company frames it as "our most capable model for agentic coding and cyber defense" and says the weights are "available to download, run, and customize." That landed roughly two weeks after the August 14 API launch — a delay Z.ai tied to a safety review of GLM-5.3's strength on offensive-security tasks, rather than the day-one weights drop it did for GLM-5.2.

For local use, the binding constraint is hardware, and Z.ai has not published a clean minimum-spec figure. GLM-5.3's 753B total parameters (40B active) are far lighter than Kimi K3's 2.8 trillion, but full-precision inference is still a multi-GPU or high-RAM-workstation job. Quantization is what closes the gap: Unsloth is producing GLM-5.3 GGUF builds — the same treatment that put a 1-bit Kimi K3 on a single Mac Studio — and the sibling GLM-5.3-Flash 3-bit GGUF already runs in 128GB of RAM. Expect a similar quantized path here, with the usual accuracy trade-off at the low-bit end.

One caveat carries over from launch. The benchmarks Z.ai published — CyberGym 84.5%, the Terminal-Bench 3.0 jump to 28.3, DeepSWE v1.1 to 66.9 — remain provider-reported. Independent labs can now rerun them against the released weights, but until they do, treat those as vendor claims rather than verified results. The Artificial Analysis Intelligence Index score of 60 is the exception: that one was measured independently, on third-party infrastructure, before the weights shipped. If you want a second open-weights option at similar measured intelligence, Qwen3.8 Max scores close behind at 58.

What people are asking

Is this actually a big deal, or just one number matching another? It's a genuine milestone with a real caveat. Tying the prior open-weights leader on an independent nine-evaluation composite, without a new pretraining run, is a meaningful post-training result. But "ties Kimi K3" as a headline glosses over the sub-score split — GLM-5.3 wins agentic work, Kimi K3 wins factual reliability — and the fact that GLM-5.3's weights aren't downloadable yet while Kimi K3's have been for weeks.

Why does Z.ai keep gating weights behind security review? Z.ai's own framing for GLM-5.3 leaned into cyber-defense capability — see explainx.ai's coverage of its CyberGym validation and the system prompt leak that surfaced separately. A model that's strong at finding vulnerabilities is also plausibly strong at exploiting them, which is a defensible reason to stage release rather than ship weights day one.

Does the price difference matter more than the tied score? For most teams, probably yes. GLM-5.3's $1.40/$4.40 per-million-token API pricing undercuts Kimi K3's $3.00/$15.00 by a wide margin at identical measured intelligence — a bigger practical lever than three points on any index, in either direction.

How reliable is a single benchmark index for choosing a model? Treat it as a starting filter, not a verdict. The Intelligence Index doesn't measure cost, latency, safety posture, or performance on your specific codebase — explainx.ai's guide to choosing open-weight vs. closed AI models covers the fuller decision framework, including licensing and deployment constraints an aggregate score never captures.

Honest limitations of the "60" score

The Intelligence Index is a relative, point-in-time snapshot, not a fixed absolute measure — Artificial Analysis periodically adds or reweights the underlying evaluations, so scores from different months aren't perfectly comparable. It also doesn't measure cost efficiency, inference latency, safety alignment, or how a model performs on a reader's own domain-specific tasks — all things that matter more than an aggregate number when actually picking a model to build on. With GLM-5.3's weights released on August 28, 2026, the "tie" now describes two models a reader can download and compare head-to-head — but GLM-5.3's cyber-defense and terminal-coding numbers stay provider-reported until independent labs rerun them, so the head-to-head is only partly settled.

Update — September 5, 2026: Artificial Analysis moved the Intelligence Index itself to v4.2 the following month — two new evaluations added, GPQA Diamond retired, and private held-out test weighting doubled to 40% to reduce benchmark gaming. The "60" score in this post reflects the version live at the time; see Artificial Analysis Intelligence Index v4.2: What Actually Changed for the current methodology and leaderboard.

Related reading

• Artificial Analysis Intelligence Index v4.2: What Actually Changed
• GLM-5.3 Max "2nd Among Open Code Models"? Fact-Checked — checking a trending aggregator headline against this same Intelligence Index and two other leaderboards
• GLM-5.3 Launch: Cyber Defense Benchmarks
• GLM-5.3 Open Weights: The Delay and the August 28 Release
• Unsloth's GLM-5.3-Flash 3-Bit GGUF Runs in 128GB RAM
• GLM-5.3 CyberGym 84.5% Independent Validation
• Kimi K3 Open Weights Are Live — 2.8T Parameters
• Kimi K3 1-Bit GGUF: Run It on a Mac Studio
• Qwen3.8 Max Open Weights Live on Hugging Face
• Complete Guide to AI Benchmarks (2026)
• Choose Open-Weight vs. Closed AI Models
• Artificial Analysis · Z.ai

Benchmark scores, pricing, and weights-release timing reflect Artificial Analysis's August 18, 2026 evaluation and Z.ai's public statements as of this post's publish date — check the linked sources directly for any updates since.