Model

GLM-5

OPEN

Zhipu AI / Z.ai (Tsinghua) · GLM-5 family · released Feb 11, 2026

Major-version flagship succeeding GLM-4.6; first Z.ai model released after Zhipu's Jan 2026 Hong Kong IPO; widely reported as trained entirely on domestic Huawei chips, no Nvidia used.

ReasoningCodingVisionFunction callingTool useAgentic

2096.8

Elo · rank #69

Parameters

744B

Active params

40B (MoE)

Context

200K tokens

Architecture

Mixture-of-Experts, 744B total / 40B active, DeepSeek Sparse Attention (DSA), 200K context, trained on a 100k-GPU Huawei Ascend 910B cluster (no Nvidia/AMD/Intel silicon)

License

MIT

Languages

—

API price (in/out)

$1 / $3.2

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

AIMEMath92.7%#34

best: GPT-5.2 · 100.0%

BrowseCompAgents62.0%#29

best: Kimi K3 · 91.2%

CyberGymCoding43.2%#13

best: GLM-5.3 · 84.5%

GPQA DiamondReasoning86.0%#58

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning30.5%#51

best: Claude Opus 5 · 64.7%

IFBenchReasoning72.0%#28

best: MiniMax M3 · 83.0%

LiveCodeBenchCoding81.9%#42

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MCP AtlasAgents67.8%#6

best: Kimi K3 · 84.2%

SWE-bench VerifiedCoding77.8%#29

best: Claude Opus 5 · 96.0%

τ²-Bench TelecomAgents89.7%#21

best: Claude Opus 4.6 · 99.3%

Terminal-Bench 2.0Coding56.2%#52

best: Gemini 3.8 Flash · 89.4%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Fine-tune it

Permissive

QLoRA505.9 GB4× H200 141GB

LoRA1584.7 GBbeyond 8× B200

Full fine-tune11941.2 GBbeyond 8× B200

QLoRA SFT on ~10k samples ≈ $20.69 (4× H200 141GB)

Plan a fine-tune →

GLM-5 family

Elo progression across releases

GLM-5Feb 20262096.8 GLM-5.1Apr 20262224.1 GLM-5.2 (High)Jun 20261000.0 GLM-5.2 (Max)Jun 20262425.7 GLM-5.3Aug 20263029.9 GLM-5.3-FlashAug 20263017.6

Resources

HFzai-org/GLM-5 DOCShttps://docs.z.ai/guides/llm/glm-5 APIAvailable via Zhipu AI / Z.ai (Tsinghua) API

API price $1/$3.2 · each benchmark row carries its own source badge (see methodology)