Model

GLM-5.2 (Max)

OPEN

Zhipu AI / Z.ai (Tsinghua) · GLM-5 family · released Jun 13, 2026

Highest-capability reasoning-effort tier: raises the thinking-token budget (~85k tokens on hard tasks vs High's ~50% of that) for complex, long-horizon agentic problems. Same weights as the High tier.

ReasoningCodingVisionFunction callingTool useAgentic

2399.0

Elo · rank #49

Parameters

744B

Active params

40B (MoE)

Context

1M tokens

Architecture

Mixture-of-Experts, 744B total / 40B active, DSA + 'IndexShare' sparse-attention indexer reuse (2.9x fewer per-token FLOPs at 1M context), 1M-token context; selectable High/Max reasoning-effort modes

License

MIT

Languages

—

API price (in/out)

$1.4 / $4.4

Modalities

text

Benchmark results

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

Agents' Last ExamAgents40.6%#13

best: GPT-6 Astra · 59.3%

AIMEMath99.2%#5

best: GPT-5.2 · 100.0%

Arena EloHuman preference1475#20

best: Claude Fable 5 · 1504

ExploitBenchAgents24.4%#11

best: GPT-6 Astra · 100.0%

FrontierBenchAgents5.1%#9

best: Claude Opus 5 · 43.3%

GDPval-AAAgents1524#20

best: Claude Fable 5 · 1932

GPQA DiamondReasoning91.2%#25

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning40.5%#31

best: Claude Opus 5.5 · 67.7%

IFBenchReasoning73.3%#25

best: MiniMax M3 · 83.0%

LiveCodeBenchCoding69.5%#74

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

MCP AtlasAgents76.8%#3

best: Step 5 Preview · 85.6%

SWE-bench ProCoding62.1%#15

best: Claude Opus 5.5 · 89.9%

SWE-bench VerifiedCoding82.8%#16

best: Claude Opus 5 · 97.0%

τ²-Bench TelecomAgents99.1%#3

best: Claude Opus 4.6 · 99.3%

Terminal-Bench 2.0Coding81.0%#20

best: DeepSeek-V4.1-Flash · 90.6%

Terminal-Bench 3.0Agents4.6%#9

best: GPT-5.6 Sol · 34.6%

Toolathlon (Verified)Agents59.9%#12

best: GLM-5.3-Flash · 78.4%

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 20262036.8 GLM-5.1Apr 20262199.4 GLM-5.2 (High)Jun 20261000.0 GLM-5.2 (Max)Jun 20262399.0 GLM-5.3Aug 20262796.9 GLM-5.3-FlashAug 20262771.9

Resources

HFzai-org/GLM-5.2 DOCShttps://docs.z.ai/guides/llm/glm-5.2 APIAvailable via Zhipu AI / Z.ai (Tsinghua) API

API price $1.4/$4.4 · each benchmark row carries its own source badge (see methodology)