Is Kimi K3 or GLM-5.2 the better open-weight model?

Different winners. Kimi K3 has the higher ceiling — Artificial Analysis Index ~57 and #1 on the Arena.ai Frontend Code Arena ahead of Claude Fable 5. GLM-5.2 is the better practical choice — 744B with a true MIT license, self-hostable today, faster to serve, and ~1/3 the output cost ($4.40 vs $15 per MTok). For raw capability pick K3; for deployable, licensed, affordable open weights pick GLM-5.2.

How much cheaper is GLM-5.2 than Kimi K3?

Substantially. GLM-5.2 costs $1.40 input / $4.40 output per MTok on the official API; Kimi K3 is $3 input / $15 output. That makes K3's output roughly 3.4x more expensive and input ~2x. For self-hosting, GLM-5.2 (744B) is far cheaper to serve than Kimi K3 (2.8T), which needs a multi-node H100 cluster.

Which open-weight model has the better license?

GLM-5.2 ships under a clean MIT license, the most permissive option for commercial use and fine-tuning. Kimi K3's weights (releasing July 27, 2026) come under a 'Modified MIT' license, which may carry additional conditions — read the terms before commercial deployment. For license clarity, GLM-5.2 is the safer pick.

Quick Answer

GLM-5.2 vs Kimi K3: Open-Weight King After K3

Published: July 26, 2026

The Short Answer

Kimi K3’s weights drop July 27, 2026 amid claims it’s the new open-weight king. The honest answer: it depends what “king” means.

Kimi K3 — the capability king. Artificial Analysis Index ~57, #1 on the Arena.ai Frontend Code Arena (ahead of Claude Fable 5). But 2.8T params, $3/$15 API, “Modified MIT” license.
GLM-5.2 — the deployment king. 744B, clean MIT license, self-hostable today, faster serving, $1.40/$4.40 (~1/3 of K3’s output cost).

If you’re paying API rates and want the top quality → Kimi K3. If you want open weights you can actually run, license, and afford → GLM-5.2.

The Comparison Table (July 2026)

	GLM-5.2	Kimi K3
Developer	Zhipu / Z.ai	Moonshot AI
Total params	744B	2.8T
Context	1M	1M
License	MIT	Modified MIT
Weights	Available now	July 27, 2026
$/MTok input	$1.40	$3.00
$/MTok output	$4.40	$15.00
AA Intelligence Index	~51	~57
Frontend Code Arena	Strong	#1 (ahead of Fable 5)
Self-host difficulty	Moderate	Very high (multi-node)
Serving speed	Faster	Slower

Where Each Wins

Kimi K3 — the ceiling. The Artificial Analysis Index ~57 and Frontend Code Arena #1 make it the strongest open model on raw capability in July 2026, beating Claude Fable 5 on frontend code. It’s the pick when quality is the only thing that matters and you’re paying API rates anyway. The costs: $3/$15 (5-17x GLM-5.2’s output on some comparisons) and a 2.8T weight footprint that needs ~16x H100 to self-host at frontier throughput.

GLM-5.2 — the practical king. 744B with a true MIT license makes it the model you can self-host, fine-tune, and ship commercially with zero license friction. It tops the open-weights Intelligence Index, serves faster than K3, and costs ~1/3 as much on output. For any team that actually intends to run open weights rather than call an API, GLM-5.2 is the default.

Does K3 Dethrone GLM-5.2?

On benchmarks and frontend coding — yes, K3 takes the capability lead. On the metric most self-hosters care about — cost-to-deploy — no, GLM-5.2 stays king. K3 is 5-17x the output price, harder to serve, and its “Modified MIT” terms need reading before commercial use. GLM-5.2 remains the sensible default; K3 is the reach model.

How to Choose

Want the highest open-weight quality, API budget → Kimi K3
Want to self-host on realistic hardware → GLM-5.2
Need clean commercial license → GLM-5.2 (MIT)
Frontend/agentic coding at the top end → Kimi K3
Cost-sensitive, high volume → GLM-5.2 (or DeepSeek V4 Pro for even cheaper)

Sources

Kimi K3 vs DeepSeek V4 Pro vs GLM-5.2 (MarkTechPost, July 18, 2026): marktechpost.com
Kimi K3 and China’s open-weight wave (DEV, July 2026): dev.to/smakosh/kimi-k3-and-chinas-open-weight-model-wave-bpp
GLM 5.2 pricing & benchmarks (OpenRouter): openrouter.ai/z-ai/glm-5.2
Kimi K3 quickstart (Moonshot AI): platform.kimi.ai/docs/guide/kimi-k3-quickstart