AI agents · OpenClaw · self-hosting · automation

Quick Answer

DeepSeek V4 vs GLM-5.2 vs Kimi K3: Which to Deploy

Published:

The Short Answer

Three Chinese labs hold the open-weight frontier in July 2026, and they optimize for different things:

  • GLM-5.2 (Zhipu/Z.ai) — best deployable model. 744B, MIT license, self-hostable today, tops the open-weights Intelligence Index (~51), cheapest to serve yourself.
  • DeepSeek V4 Pro (DeepSeek) — cheapest API. $0.435/$0.87 off-peak, leads raw SWE-bench Verified and competitive coding.
  • Kimi K3 (Moonshot AI) — highest ceiling. 2.8T, Artificial Analysis Index ~57, leads the Frontend Code Arena — but $3/$15 API and a multi-node cluster to self-host.

If you’re choosing one to actually deploy: GLM-5.2. It’s the only one of the three that self-hosts on realistic hardware while still being frontier-competitive.

The Comparison Table (July 2026)

DeepSeek V4 ProGLM-5.2Kimi K3
DeveloperDeepSeekZhipu / Z.aiMoonshot AI
Total params1.6T (49B active)744B2.8T
Context1M1M1M
LicenseOpen weightsMITModified MIT
Weights availableYesYes (today)July 27, 2026
API input $/MTok$0.435 (off-peak)$1.40$3.00
API output $/MTok$0.87 (off-peak)$4.40$15.00
AA Intelligence Index~mid~51 (top open)~57
Coding strengthSWE-bench / competitiveAgentic / long-contextFrontend Code Arena
Self-host difficultyVery high (data-center)ModerateVery high (multi-node)

DeepSeek V4 Pro pricing doubles during peak hours (1-4 and 6-10 UTC). Cache-hit input drops to ~$0.043.

Where Each One Wins

DeepSeek V4 Pro — the cost floor. At $0.435/$0.87 off-peak it anchors the entire market’s pricing floor. It leads raw SWE-bench Verified, LiveCodeBench (~93.5%), and Codeforces (~3206). If your workload is competitive coding or high-volume agent loops and you can schedule around peak hours, nothing beats its cost-per-task. The catch: peak/valley surge pricing (2x during CN business hours) and the deepseek-chat/deepseek-reasoner legacy names were retired July 24, 2026 — migrate to deepseek-v4-pro / deepseek-v4-flash.

GLM-5.2 — the deployable king. 744B with an actual MIT license makes it the model you can self-host, fine-tune, and ship commercially without lawyer meetings. It tops the open-weights Intelligence Index and is faster than both rivals at serving time. At $1.40/$4.40 it’s ~1/3 the output cost of Kimi K3. This is the default recommendation for teams that want open weights they can run.

Kimi K3 — the capability ceiling. The Artificial Analysis Index ~57 and Frontend Code Arena lead (ahead of Claude Fable 5) make it the strongest open model on raw capability. But $3/$15 API pricing is 5-17x GLM-5.2’s output rate, and self-hosting the 2.8T weights (dropping July 27) needs a ~16x H100 cluster. Choose K3 when you need the highest open-weight quality and are paying API rates anyway.

How to Choose

  • Want to self-host on real hardware? → GLM-5.2
  • Cheapest API, can schedule around peak hours? → DeepSeek V4 Pro
  • Highest open-weight quality, API budget? → Kimi K3
  • Competitive/algorithmic coding? → DeepSeek V4 Pro
  • Frontend/agentic coding at the top end? → Kimi K3
  • Long-context knowledge work, commercial license clarity? → GLM-5.2 (MIT)

Sources