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.
Updated 2026-09-08
Key points
AA Intelligence Index (tied open-source lead)
Tied for first among open models, just 1 point behind the closed flagship Sol (61).
Total parameters
GLM-5.3 reuses 5.2's 753B base (40B active); K3 is a 2.8T Stable LatentMoE (896 experts, 16 active).
GLM-5.3 official price
Per million tokens input/output. K3's official price is $3/$15 — GLM is less than half the price.
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
Kimi K3 launches (WAIC); weights go open 07-27.
GLM-5.3 launches: same base + post-training scaling, coding +50%, topping CyberGym at 84.5%.
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. QCode catalog prices are set independently and do not track vendor list prices line for line: glm-5.3 is currently listed at $2.38 in / $8.32 out / $0.59 cache read per million tokens.
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.
Related reading
GLM-5.3 guide
The details of post-training scaling.
Kimi K3 guide
Calling the 2.8T and its reasoning tiers.
GLM-5.3 vs GLM-5.2
How much the same base improved.
Not affiliated with Zhipu / Moonshot. Leaderboard numbers are a 2026-08 snapshot.