Model

GLM-4.6

OPEN

Zhipu AI / Z.ai (Tsinghua) · GLM-4.6 family · released Sep 30, 2025

Flagship coding/agentic update to GLM-4.5; expanded context to 200K, ~15% more token-efficient, near-parity with Claude Sonnet 4 on real-world coding (CC-Bench).

ReasoningCodingVisionFunction callingTool useAgentic

1899.3

Elo · rank #83

Parameters

357B

Active params

32B (MoE)

Context

200K tokens

Architecture

Mixture-of-Experts, 357B total / 32B active, 200K context (up from 128K in GLM-4.5), hybrid thinking/non-thinking response modes

License

MIT

Languages

—

API price (in/out)

$0.6 / $2.2

Modalities

text

Benchmark results

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

AIMEMath93.9%#25

best: GPT-5.2 · 100.0%

BrowseCompAgents45.1%#35

best: Kimi K3 · 91.2%

GPQA DiamondReasoning81.0%#82

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning17.2%#76

best: Claude Opus 5 · 64.7%

IFBenchReasoning36.7%#64

best: MiniMax M3 · 83.0%

LiveCodeBenchCoding82.8%#38

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

SWE-bench VerifiedCoding68.0%#68

best: Claude Opus 5 · 96.0%

τ²-Bench TelecomAgents75.9%#25

best: Claude Opus 4.6 · 99.3%

Terminal-BenchCoding40.5%#7

best: Claude Opus 4.5 (High) · 59.3%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Fine-tune it

Permissive

QLoRA242.8 GB2× H200 141GB

LoRA760.4 GB4× B200 192GB

Full fine-tune5729.9 GBbeyond 8× B200

QLoRA SFT on ~10k samples ≈ $16.55 (2× H200 141GB)

Plan a fine-tune →

GLM-4.6 family

Elo progression across releases

GLM-4.6Sep 20251899.3

Resources

HFzai-org/GLM-4.6 DOCShttps://docs.z.ai/guides/llm/glm-4.6 APIAvailable via Zhipu AI / Z.ai (Tsinghua) API

API price $0.6/$2.2 · each benchmark row carries its own source badge (see methodology)