PyyanThe AI Index

Pyyan / Compare / Phi-4 vs Kimi K3 vs GLM-5.2 vs Qwen3-235B-A22B

Phi-4 vs Kimi K3 vs GLM-5.2 vs Qwen3-235B-A22B

4 of 5

Open-Weight Models · verified 13 Aug 2026

×Phi-4Microsoftcurrent
×Kimi K3Moonshot AIcurrent
×GLM-5.2Zhipu AIcurrent
×Qwen3-235B-A22BAlibabacurrent
1 slot left
SpecificationPhi-4Kimi K3GLM-5.2Qwen3-235B-A22B
SummaryBest small model — fits a 12GB GPU.Largest open-weight model shipped.Strongest open reasoning and mathematics.Cheapest capable model, most permissive licence.
LicenceMITOpen weightsOpen weightsApache-2.0
Parameters14B2.8T—235B MoE
Released—17 Jul 2026——
CategoryOpen-Weight ModelsOpen-Weight ModelsOpen-Weight ModelsOpen-Weight Models
OfficialMicrosoft ↗Moonshot AI ↗Zhipu AI ↗Alibaba ↗

Highlighted rows are where these differ.

Phi-4

  • Beats GPT-4o on mathematics
  • Phi-4-mini runs in ~3GB at Q4

Best for local use.

Full spec sheet →

Kimi K3

  • Tops the Frontend Code Arena
  • 76.8% on SWE-Bench Verified

Best for coding.

Full spec sheet →

GLM-5.2

  • 91.2 on GPQA Diamond
  • Strong agentic tool use

Best for maths and reasoning.

Full spec sheet →

Qwen3-235B-A22B

  • No licence conditions whatsoever
  • 22B active of 235B total

Best for commercial self-hosting.

Full spec sheet →