Z.ai released GLM-5.2 on June 13, 2026, rolling it out first to GLM Coding Plan members. The MIT-licensed weights and a technical blog followed on June 16.

The 753-billion-parameter mixture-of-experts model uses Dense-Sparse-Alternating attention with IndexShare. This cuts per-token FLOPs by 2.9x at 1M context. It supports tool calling, structured output, reasoning traces, and context caching natively.

Benchmarks published by NVIDIA's NGC catalog place GLM-5.2 within a percentage point of Anthropic's Opus 4.8 on agentic tests. MCP-Atlas scores 76.8 versus 71.8 for GLM-5.1. Tool-Decathlon reaches 48.2 versus 40.7. SWE-bench Pro hits 62.1. Terminal Bench 2.1 reaches 81.0.

reported enterprises are seeing roughly a fifth of the token cost compared to Opus 4.8 for comparable agentic workloads. OpenRouter traffic climbed faster than after DeepSeek V4's April launch.

Confirmed

  • Model: GLM-5.2, 753B parameters, MoE with Dense-Sparse-Alternating and IndexShare attention
  • License: MIT — commercial use, modification, fine-tuning, distillation, and self-hosting permitted without royalty or Z.ai usage restrictions
  • Context window: 1,000,000 tokens input and output
  • Release timeline: June 13 (Coding Plan early access), June 16 (public weights on Hugging Face, technical blog, MIT license)
  • Key benchmarks (vs. GLM-5.1): HLE 40.5→31.0, HLE w/ Tools 54.7→52.3, AIME 2026 99.2→95.3, GPQA-Diamond 91.2→86.2, SWE-bench Pro 62.1→58.4, NL2Repo 48.9→42.7, Terminal Bench 2.1 81.0→63.5, MCP-Atlas 76.8→71.8, Tool-Decathlon 48.2→40.7
  • Inference stack: SGLang (v0.5.13+), vLLM (v0.23.0+), KTransformers (v0.5.12+), Transformers (v0.5.12+); tested on NVIDIA Grace Blackwell GB200/GB300
  • Deployment: Available via NVIDIA NIM, OpenRouter, Fireworks, Together, and self-hosted; OpenAI-compatible endpoint drops into Cursor, Claude Code, Cline, Roo Code, and 20+ tools
  • Cost comparison (hosted): ~$0.50–$2 / 1M input, ~$1–$4 / 1M output vs. Opus 4 at ~$15 / ~$75; self-hosted infrastructure cost only
  • Real-world test: 45-minute autonomous bug-hunting session in Claude Code processed 6M tokens for $3.36 (72% cache rate); surfaced 20 Sentry errors, 5 Vercel signals, 14 planned fixes including 2 P0s; HTML/CSS reliable, React under multi-step pressure shakier
  • Design taste: Beats Claude Opus on Fable 5 benchmark for UI/UX critique, HTML/CSS aesthetic generation, and design vocabulary precision
  • Geopolitical context: Release coincided with U.S. government restrictions on Anthropic's Fable and OpenAI's GPT-5.6, making a non-revocable open-weight model strategically attractive for enterprises

Analysis

GLM-5.2 represents a genuine threshold crossing for open-weight models. It is the first that feels production-ready inside coding harnesses like Claude Code and Cursor without constant babysitting. The combination of MIT licensing, 1M context, and agentic benchmark parity with Opus 4.8 changes the procurement calculus for teams accumulating long-context agentic sessions — exactly where frontier API costs compound fastest.

Nathan Lambert at Interconnects notes the 204-day gap (6.8 months) between Opus 4.5 (Nov 2025) and GLM-5.2 (June 2026). This sits squarely in the 6–9 month lag pattern often cited between U.S. closed labs and Chinese open counterparts.

The cost structure is the lever. Hosted inference at 80–90% below Opus 4.8 per token, plus self-hosting optionality, means enterprises can run sustained agentic workflows without the token budget anxiety that currently gates adoption. Multi-hour coding agents, log analysis, and autonomous testing become viable. The live test ($3.36 for a 45-minute, 6M-token session) is a concrete proof point: the economics work at the workload level, not just the per-token level.

Limitations are real and documented. React-heavy codebases under multi-step agentic pressure remain a friction point; the model recovers but requires verification. The 753B parameter count demands serious GPU infrastructure for self-hosting (GB200/GB300 class), putting true self-hosted sovereignty out of reach for smaller teams unless they rent inference. Training data, evaluation methodology, and data collection details remain undisclosed.

Unknown

  • Exact training dataset composition, size, and collection methodology (listed as "Undisclosed" in NVIDIA model card)
  • Detailed evaluation dataset construction and labeling processes for the reported benchmarks
  • GLM-5.5 rumors — no official confirmation, timeline, or specs from Z.ai; this article covers only the shipped 5.2 release
  • Long-term maintenance commitment: whether Z.ai will publish iterative patches, quantized variants, or a 5.3 follow-up on a predictable cadence
  • Enterprise support SLAs, indemnification, or commercial redistribution terms beyond the MIT license

Our take

GLM-5.2 earns a slot in the daily rotation for production agentic coding — not as a stunt, but because harness ergonomics, context length, and per-token economics align. The React weakness is a genuine gating factor; test your specific stack before committing critical paths. The strategic shift is vendor leverage: an MIT-licensed model that runs on any compliant inference provider means the next API price hike or policy change is a negotiation, not a crisis.

Series: 1. GLM-5.2: What Actually Shipped and Why It Matters · 2. The 2026 Roadmap: Zhipu's Release Cadence and Strategy · 3. GLM-5.5 Rumors vs. Official Silence: Parsing the August 2026 Reports · 4. Developer's Guide: What to Treat as Confirmed vs. Unknown in the GLM Ecosystem · 5. GLM-5.3 Ships: Post-Training Coding and Emergent Cyber Defense · Zhipu GLM

Sources