Open-source duel · 2026-08-22

GLM-5.3 vs Kimi K3
A family fight at the top of open source

The two are tied for the open-source lead on the Artificial Analysis Intelligence Index (~60). GLM-5.3 got there with post-training scaling, K3 with a 2.8T base — completely different routes, and different prices.

#glm-5.3#kimi k3#open-source leaders#comparison

Key points

≈60

AA Intelligence Index (tied open-source lead)

Tied for first among open models, just 1 point behind the closed flagship Sol (61).

753B / 2.8T

Total parameters

GLM-5.3 reuses 5.2's 753B base (40B active); K3 is a 2.8T Stable LatentMoE (896 experts, 16 active).

$1.40 / $4.40

GLM-5.3 official price

Per million tokens input/output. K3's official price is $3/$15 — GLM is less than half the price.

84.5%

GLM-5.3 CyberGym

Tops the vulnerability-discovery benchmark above Mythos 5 (83.8) and Sol (83.6) — the first open model to lead that board.

Two different technical routes

GLM-5.3 (08-14) didn't change the base: it's the same 753B MoE as 5.2, with all the gains coming from 'extreme post-training scaling' — the official claim is a 50% better coding feel, with coding + cyber defense as the focus. Kimi K3 (07-16) is brute force done right: 2.8T, the largest open weights ever, with mandatory reasoning (low/high/max). One proves the value of post-training; the other proves scale still works.

Open source's August

The open-weights camp has been shipping nonstop this month: GLM-5.3 (released 08-14, API live 08-19), K3's weights going open (scaling up since 07-27), Qwen3.8-Max opening Max-tier weights for the first time (08-12), and the DeepSeek V4 line iterating. Hugging Face data shows Chinese open models at 41% of downloads, passing the US for the first time — open source is no longer the 'budget alternative.'

Timeline

2026-07-16

Kimi K3 launches (WAIC); weights go open 07-27.

2026-08-14

GLM-5.3 launches: same base + post-training scaling, coding +50%, topping CyberGym at 84.5%.

2026-08-19

GLM-5.3 API goes live; the two tie for the AA open-source lead.

Confirmed vs watch out

Confirmed

Both models' release dates, parameters, prices and leaderboard positions are backed by official pages and re-checkable third parties; both have real call volume on QCode.

Watch out

'Tied for the lead' is specifically the AA Intelligence Index; the order changes on other boards (K3 is stronger on the frontend Code Arena, GLM-5.3 stronger on CyberGym). Read the per-task breakdowns, not the total score.

Pick by scenario

Pick GLM-5.3

Coding + security-review tasks, budget-sensitive ($1.40/$4.40), 1M context. Reasoning is adjustable across three tiers.

Pick Kimi K3

Ultra-long context plus heavy reasoning, or when you need the largest open base. $3/$15, and mind the token volume from mandatory reasoning.

How to verify yourself

Both have open weights and both have APIs. Go straight to API testing: run the same task batch on both and convert by 'first-pass rate × total tokens' — GLM-5.3's unit price is lower, K3's mandatory reasoning burns more tokens; the converted number is your answer.

On QCode

Both glm-5.3 and kimi-k3 are on sale (with real calls in the last 30 days), switchable with the same key at official price times the service rate. The QCode catalog currently lists glm-5.3 at $1.40/$4.40.

FAQ

Which is stronger, GLM-5.3 or Kimi K3?

They tie for the open-source lead on the AA Intelligence Index (~60). Each wins its own categories: GLM-5.3 leads in coding and security (CyberGym 84.5%, first place), K3 leads in frontend (Code Arena 1679 Elo).

How big is the price gap?

Official API: GLM-5.3 at $1.40/$4.40, K3 at $3/$15. GLM is less than half the price; but K3 has mandatory reasoning while GLM's is adjustable — convert by task.

Are both open-source?

K3's weights went open 07-27 (2.8T, the largest ever). GLM-5.3's team says weights will open in stages subject to safety review — follow official channels.

How big are the context windows?

Both are 1M-token class. GLM-5.3's max output is 128K; K3 is positioned for long-context office work.

How far behind the closed flagships are they?

AA index: both around 60, Sol at 61 — a 1-point gap. As of 2026-08 the capability gap between open and closed flagships is basically closed.

Can I call both on QCode?

Both are on sale, switchable with the same key. Both have real call volume in the last 30 days.

Sources

Zhipu's official release (08-14) and API launch announcement (08-19), Moonshot official (07-16/07-27), Artificial Analysis, Caixin (08-14), SegmentFault third-party review (08-20), QCode /models (08-22).

The open-source duo, tested with one key

glm-5.3 and kimi-k3 are both on sale on QCode — settle it with real tasks.