Large Language Models

GLM-5.2

Z.ai's open-weight long-horizon flagship — a 753B-param MoE with a stable 1M context under an MIT license, beating GPT-5.5 on several coding benchmarks.

GLM-5.2, released June 13 2026 by Z.ai (formerly Zhipu AI), is the year’s open-weight leader for long-horizon coding. It’s a 753-billion-parameter mixture-of-experts (~40B active) and, for the first time in the GLM line, delivers its capability on a stable 1M-token context — enabled with the glm-5.2[1m] identifier — all under a pure MIT license with no regional limits.

Its new IndexShare architecture reuses one indexer across every four sparse-attention layers, cutting per-token FLOPs ~2.9× at 1M context, which is what makes that window economical. On benchmarks it tops the open field and edges several closed rivals: SWE-bench Pro 62.1 (vs GPT-5.5’s 58.6 and GLM-5.1’s 58.4), FrontierSWE 74.4% (near Claude Opus 4.8’s 75.1%), and AIME 2026 99.2. Coding-plan subscriptions start at just $12.60/month, with API at ~$1.40 / $4.40 per 1M tokens.

The trade-offs are deployment and tuning: a 753B MoE realistically needs a multi-GPU node to self-host, and the High/Max reasoning-effort levels trade latency and quota for quality. For benchmark-leading open coding, it’s at the front of the pack.

Best for

  • Long-horizon autonomous coding and engineering
  • Repo-wide reasoning at 1M context
  • Agentic dev tools (Claude Code, Aider, Cline)

Pros & cons

Strengths

  • Leads the open cluster — SWE-bench Pro 62.1, beating GPT-5.5 (58.6)
  • Stable 1M-token context via the IndexShare architecture
  • Permissive MIT license with no regional limits

Limitations

  • A 753B-param MoE needs a multi-GPU node to self-host
  • High/Max reasoning effort raises latency and quota cost

Sources

Last updated: 2026-06-18 · Specs and pricing change fast — verify on the vendor's site before relying on them.