GLM-5.3 vs Kimi K3: 753B vs 2.8T. The Older Weights Just Won on Points.

Zhipu’s post-training special vs Moonshot’s trillion-parameter giant — the first real head-to-head of China’s two open-weight flagships, with receipts and asterisks.

GLM-5.3 launched on August 14, 2026, the day I’m writing this. Its numbers are vendor-reported (v) and hours old. Kimi K3’s numbers are a mix of vendor-reported (v) and independently verified (i) figures, after three weeks of third-party scrutiny. Where a comparison is clean, I say so. Where it isn’t, I say that instead.

Two flagships, two opposite theories of progress

Three weeks ago, Moonshot AI shipped Kimi K3 and made the biggest bet in open-weight AI this year: scale. A 2.8-trillion-parameter mixture-of-experts model. 104B active per token, 896 routed experts, a hybrid linear-attention architecture nobody else runs, native vision, a million-token context. The first open model in the 3T class. The implicit claim: to reach the frontier, you build a bigger brain.

Yesterday, Zhipu shipped GLM-5.3 and made the opposite bet: post-training. The base model is byte-identical to June’s GLM-5.2. Same ~753B parameters, same architecture, same weights. Every claimed gain (and the headline claim is a 50% coding improvement over its predecessor) comes from reinforcement learning applied to a frozen base. Zhipu’s own footnote is the thesis: "We may still be far from reaching the intelligence ceiling of this base model."

Right now these are the two best open-weight coding models in the world, and they happen to be built on the two competing answers to the only question that matters in open AI: does the next capability jump come from pretraining compute, or from RL?

So here’s the full comparison. Architecture, benchmarks, cost, behavior, licensing, and the fine print both launch decks would rather you skipped.

The Contenders

GLM-5.3 Kimi K3
Lab Zhipu AI (Z.ai), Beijing Moonshot AI, Beijing
Released Aug 14, 2026 (today) Jul 22–23, 2026
Architecture MoE, ~753B total / ~40B active MoE, 2.78T total / 104B active
What’s new Post-training only; same base as GLM-5.2 Everything: new KDA hybrid attention, 93 layers, 896 experts
Context 1M tokens (presumed; unconfirmed) 1M tokens (confirmed)
Vision ❌ Text-only ✅ Native (401M-param ViT)
Weights Promised in ~2 weeks, license TBD (5.2 was MIT) Public since Jul 27 — custom Kimi K3 License, not MIT
API price Unannounced; GLM-5.2 ran $1.40 in / $4.40 out $3.00 in / $15.00 out ($0.30 cache-hit)
Headline claim "Strongest open coding model; approaching Claude Fable 5" "Open frontier intelligence," while candidly admitting it trails Fable 5 and GPT-5.6 Sol

Two rows of that table deserve emphasis.

First, K3 is 3.7× larger in total parameters and ~2.6× larger in active parameters. Whatever GLM-5.3 achieves, it achieves with roughly a quarter of the brain. That matters for self-hosting economics more than for API users, but it matters.

Second, the openness asymmetry is real and current: you can download K3 from Hugging Face today. You cannot download GLM-5.3 at all. Zhipu is holding weights for two weeks of safety evaluation and hardening — the same staged-openness play GLM-5.2 ran in June, except longer, and with a more defensible excuse this time (the model’s offensive-security capability jumped; more on that below). Today, "open-weight GLM-5.3" does not exist. K3 wins this row by forfeit.

The head-to-head: only five benchmarks overlap cleanly

Here’s the discipline problem with this comparison. Both labs published big tables, but they mostly published different benchmarks, and where they overlap, harness choices muddy the water. K3’s vendor table was produced largely under its native Kimi Code CLI; GLM-5.3’s numbers come from Zhipu’s own harnesses. Harness swaps can move agentic scores by 10+ points, as Moonshot’s own footnotes demonstrate.

Filtering for benchmarks where both models have published numbers on the same task set:

Benchmark GLM-5.3 Kimi K3 Notes
DeepSWE v1.1 (long-horizon SWE) 66.9 (v) 67.5 (v) Statistical tie; K3 also posts 67.3 on the neutral mini-SWE-agent harness
Agents’ Last Exam 28.5 (v) 28.3 (i, leaderboard) Tie, different harnesses
Toolathlon-Verified 73.0 (v) 76.5 (v) K3 ahead
GDPval-AA v2 (44-occupation knowledge work, Elo) 1,769 (v) 1,686 (i) GLM-5.3 claims the lead — and above Fable 5’s 1,747 — but vendor-reported
AutomationBench 48.2 (v) 30.8 (v) Huge GLM lead; version/harness mismatch likely, treat as directional

What that table says, once you strip the noise: on the shared agentic-coding core (DeepSWE, Agents’ Last Exam), the two models are dead even. A two-trillion-parameter gap and a completely different training philosophy bought Moonshot nothing on DeepSWE. 67.5 versus 66.9 is inside the noise of harness variance. Zhipu’s post-training bet matched Moonshot’s scale bet on the benchmarks both labs bothered to run.

Where GLM-5.3 pulls ahead (GDPval, AutomationBench), the numbers are vendor-fresh and unverified. Where K3 pulls ahead (Toolathlon), its number is three weeks old and has survived third-party poking.

Edge on evidence quality: K3. Edge on trajectory: GLM-5.3, which was scoring 46.2 on DeepSWE two months ago.

The Terminal-Bench trap

You will see hot takes today comparing K3’s 88.3 on Terminal-Bench to GLM-5.3’s 28.3 on Terminal-Bench and declaring a massacre. Don’t fall for it. These are different benchmark versions.

K3’s 88.3 (vendor) and ~85 (Artificial Analysis, independent) are on Terminal-Bench 2.1. GLM-5.3’s 28.3 is on Terminal-Bench 3.0, a substantially harder 2026 revision, under which GLM-5.2 — a model that scored ~81–83 on 2.1 — collapsed to 4.6. The 3.0 leaderboard is young and sparsely populated; nobody outside Zhipu has run K3 on it yet.

The honest sentence is: GLM-5.3 improved ~6× over its predecessor on the newest, hardest terminal benchmark, and nobody knows where K3 lands on it. If you see a "28.3 vs 88.3" comparison in the wild this week, you’re reading someone who skipped the footnotes. That’s also a preview of launch-day epistemics in general: everything from Zhipu is single-source until the weights drop and the third-party harnesses spin up.

Where K3 is simply more model

Benchmarks obscure K3’s structural advantages, so let me list them plainly.

Vision. K3 sees. Natively — a 401M-parameter vision encoder trained in from the start, not a bolt-on. It posts 91.1 on OmniDocBench, 90.0 on Video-MME, and competitive numbers across the visual-reasoning suite. This is what powered its most famous independent result: #1 on LMArena’s WebDev arena, Elo 1,679, beating Claude Fable 5 in 76% of blind matchups, the first open-weight model to top a flagship arena board. Frontend work is visual work; K3’s arena crown and its vision stack are the same story.

GLM-5.3 is text-only. Zhipu keeps multimodality in the separate, closed GLM-V line, despite Jie Tang’s June 29 developer poll (466K views) returning a near-unanimous "vision" as the community’s top request. If your workload involves screenshots, PDFs, design mockups, or UI states, this comparison is over before it starts. K3 is the only option of the two.

Breadth. K3 was built to be a general frontier model: 93.5 on GPQA Diamond, 74.7 on AA-LCR long-context reasoning, 91.2 on BrowseComp deep research, 84.8 on OSWorld-Verified computer use. Its independent Artificial Analysis Intelligence Index score of 57 (i), #4 overall and #1 open, reflects that breadth. GLM-5.2, by contrast, was a specialist: AA 51, strong on coding, invisible on math and creative tasks. Nothing in GLM-5.3’s launch suggests a scope change — same base model means same shape of talent, sharpened. Zhipu published no GPQA, no arena data, no vision anything. GLM-5.3 is a coding-and-agents weapon. K3 is a generalist that happens to code at the same level.

Verifiability. K3 has been poked by Artificial Analysis ($2,710 of independent eval), Superconductor (real Rails tickets), Fireworks (a 1,030-task same-harness routing study against Fable 5), and the arena’s blind voters. Its profile — strengths, slowness, token hunger, proactiveness — is known. GLM-5.3’s profile is a press release.

Where GLM-5.3 punches back

The efficiency story. Zhipu’s internal Code Bench data shows GLM-5.3 completing 31.4% of hard tasks on ~50K output tokens, where Claude Opus 4.8 needs ~120K for 29.5%. K3, meanwhile, is documented as verbose: 130M output tokens across AA’s full evaluation suite versus a 63M median, and the slowest agent Superconductor tested at ~44 minutes per ticket. When I ran K3 against my race-condition bug last month, it fixed the bug correctly and then spent 40 unrequested minutes refactoring my idempotency layer into a middleware — a behavior Moonshot’s own docs warn about ("excessive proactiveness"). GLM-5.3’s pitch is the opposite: a steeper reasoning-effort curve (accuracy climbs from ~24.5% at Low to 34.5% at Max, where GLM-5.2’s curve was nearly flat) that converts extra tokens into extra accuracy instead of extra opinions.

Security. GLM-5.3’s most dramatic jumps are cyber: 84.5% on CyberGym, above Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%), plus a doubled 54.4% on ExploitBench, though still well behind the closed frontier there. K3’s table shows no dedicated security evals beyond the cyber-flavored tasks inside Kimi Code Bench. If you do defensive security work, GLM-5.3 is making a specific, quantified claim on your attention. K3 isn’t competing for it.

Price (probably). GLM-5.3 pricing wasn’t announced, but GLM-5.2 ran $1.40/$4.40 per million tokens against K3’s $3.00/$15.00. If 5.3 inherits anything close to that bracket, it undercuts K3 by ~2× on input and ~3.4× on output. K3’s defenders will note its Mooncake serving stack hits >90% cache-hit rates in coding workloads, making the effective input price $0.30. Fair point. But output tokens are where agentic bills live, and $4.40 vs $15.00 is not a rounding error. AA’s cost-per-task figures (K3 at $0.94 vs GLM-5.2’s $0.47) already reflected this; GLM-5.3’s token-efficiency claims would widen it.

Size. If you self-host (the entire point of open weights for many teams), ~753B/40B-active is a very different hardware conversation than 2.78T/104B-active. K3’s weights exist and GLM-5.3’s don’t, so this is theoretical for two more weeks. But when both are downloadable, GLM-5.3 will run on roughly a third of the GPUs.

The licensing fine print

This is the row most comparison pieces skip, and it may matter more than any benchmark.

K3 ships under a custom "Kimi K3 License" — not MIT, not Apache. Commercial hosting above a $20M revenue threshold requires a separate agreement with Moonshot. Several secondary sources flatly get this wrong and call it open-source MIT. It is open-weight, with strings.

GLM-5.3’s license is unknown, but GLM-5.2 shipped MIT, the most permissive option on the table, and Zhipu has framed open weights as strategic. If 5.3 lands MIT in two weeks, it becomes the more legally comfortable model for commercial self-hosting despite arriving later. If it doesn’t, K3’s string-attached-but-present weights beat Zhipu’s absent ones.

Also in K3’s risk column: the White House has publicly accused Moonshot of distilling K3 from Claude Fable 5 via restricted chips. Independent experts note the timeline doesn’t support the claim, but the accusation is now part of the model’s procurement profile whether it’s true or not. GLM carries no equivalent political baggage this month, just the standard data-residency conversation that applies to every Chinese API.

The behavior report: working with each

K3’s working style is documented and consistent across every independent tester: brilliant, thorough, slow (35 tok/s output vs a ~78 median), verbose, prone to scope creep. It is a queue-and-review model, not an iterate-while-you-watch model. It also demands harness hygiene — Moonshot warns that dropping its reasoning_content across turns makes output "highly unstable," which is why its native Kimi Code CLI is the fair way to run it.

GLM-5.3’s working style is, today, a marketing page. The effort-level mechanism (Non-Thinking → Max) promises controllable latency/accuracy trade-offs; if it works as charted, that’s a real UX advantage over K3’s max-effort-by-default launch state (K3’s low/high effort modes came post-launch). Zhipu’s /api/anthropic drop-in endpoint also means GLM-5.3 slides into Claude Code and similar harnesses without ceremony, where K3’s harness sensitivity makes it fussier off its home turf.

One practical launch-day note: Zhipu reset all GLM Coding Plan quotas the same morning, one day after DeepSeek telegraphed a significant V4 price hike. The Chinese labs are in a price-and-quota knife fight, and users are the beneficiaries. For now.

The asterisks

GLM-5.3:

  • Every number is vendor-reported, single-source, hours old. GLM-5.2’s launch numbers confirmed in direction but narrowed in scope once independent evals landed. Expect the same.
  • Weights don’t exist publicly yet. "Open-weight leader" claims are two weeks premature.
  • No vision, no published general-reasoning scores, pricing unannounced.
  • The AutomationBench and GDPval gaps over K3 are the numbers most likely to shrink under neutral harnesses.
  • The 4.6 → 28.3 Terminal-Bench 3.0 jump says as much about how under-trained 5.2 was as how strong 5.3 is — Zhipu essentially admits this.

Kimi K3:

  • The custom license with the $20M commercial clause. Read it before your legal team discovers it.
  • Slowest-in-class output speed and well-documented token hunger; the sticker price understates the invoice.
  • Hallucination regressed versus K2.6 on AA’s Omniscience eval (39% → 51% non-hallucination) even as accuracy rose.
  • Harness-sensitive: unstable output in harnesses that drop reasoning history; vendor-documented "excessive proactiveness" needs system-prompt guardrails.
  • The unresolved distillation accusation is a procurement risk regardless of its merit.

What happens next

This comparison has an expiry date, and both labs set it.

Zhipu has already said the quiet part aloud: the next major GLM version moves to a brand-new architecture with doubled parameters, explicitly targeting Fable 5 across the board. GLM-5.3, by Zhipu’s own framing, is the intermission, proof their RL stack works before they point it at a ~1.5T base. When Jie Tang was asked about GLM-6 on X this week, his answer was the same single word he posted hours before 5.3 shipped: "soon."

Moonshot, meanwhile, now has three weeks of ecosystem feedback on K3 and a technical report’s worth of headroom to iterate. A K3.1 with GLM-style post-training intensity is the obvious counterpunch, and if K3’s RL-scaling figures in its own report are any indication, Moonshot knows exactly how much capability is still sitting in its 2.8T base.

In other words, the scale-vs-post-training question won’t stay a choice for long. Both labs are converging on "both." The interesting fights are the next two checkpoints: GLM-5.3’s weight drop (~August 28) and whatever Moonshot answers with.

The honest math

Scenario Pick Why
Frontend / design-to-code / anything visual K3 Only one of these models can see; its WebDev arena #1 is independently verified
Pure terminal/agentic coding, API, today Toss-up leaning GLM-5.3 DeepSWE/ALE parity at a (likely) third of the output-token price, pending independent confirmation
Self-hosting, this week K3 It’s the only one that exists as weights
Self-hosting, after GLM-5.3’s weight drop Depends on license MIT + 753B beats custom-license + 2.8T for most teams’ legal and GPU budgets
Long-horizon batch jobs, queue-and-review K3 Documented strength; pace doesn’t matter overnight
Watching the agent work interactively GLM-5.3 (provisionally) K3’s 35 tok/s and scope creep are torture live; GLM’s effort tiers promise control
Defensive security / vuln research GLM-5.3 CyberGym 84.5% is the best published number, closed models included
Broad knowledge work beyond code K3 GPQA, BrowseComp, OSWorld, GDPval — GLM didn’t even show up to these rows

The verdict

Three weeks ago, Kimi K3 was the answer to "what’s the best open-weight model?" A 2.8T monument to scale that matched the closed frontier on coding and beat it, blind-tested, on frontend work.

GLM-5.3’s answer is uncomfortable for that thesis: Zhipu took a base model a quarter of K3’s size, changed zero weights, and reached the same DeepSWE score through reinforcement learning alone. If the vendor numbers hold (a real if, maybe the biggest if in open-weight AI this month), then the marginal trillion parameters bought Moonshot generality and vision, but not coding superiority.

My working split, pending GLM-5.3’s weight drop and independent evals: K3 stays my default for anything visual, anything general, and anything I need to self-host this month. It’s proven in a way a launch-day model cannot be. GLM-5.3 takes over my Claude Code backend for terminal-heavy agentic work the moment its numbers survive a harness that isn’t Zhipu’s, because the efficiency story, if real, is worth more than the benchmark story.

Honestly, which one "wins" matters less than what the pair of them proves. Open models now have two live routes to the frontier — build bigger, or train smarter on what you have — and in two weeks, when GLM-5.3’s weights land next to K3’s on Hugging Face, anyone with the GPUs can test both routes for free.

That’s the comparison worth running. I’ll publish the receipts when I do.

Sources

Tags: Artificial Intelligence, Machine Learning, Technology, Programming, LLM

About the author: I run the same bug against every model so you don’t have to. Next up: the GLM-5.3 weight-drop re-test, where we find out which of Zhipu’s launch numbers survive a neutral harness. More model audits in my newsletter.