GLM-5.2 Open-Sourced: 1M Context, Top-Tier Coding Model from Zhipu AI

Forum topic · QianXun · 2026-06-18

Summary

On June 16, 2026, Zhipu AI released and open-sourced GLM-5.2 under the MIT license. The model features a 1M-token context window and scored 51 on the Artificial Analysis composite leaderboard, placing it alongside Anthropic and OpenAI in the top three as open-source SOTA, while ranking first in Code Arena's million-user blind testing. Key breakthroughs include a 'solid' 1M context that maintains quality at scale—achieved by training on real long-horizon coding agent scenarios—and infrastructure advances: IndexShare sparse attention (reusing one indexer every four layers, cutting per-token FLOPs to 2.9x), MTP speculative decoding with up to 20% longer acceptance, and the in-house Slime framework for agentic RL. GLM-5.2 ran day-one on nine domestic Chinese compute platforms (Ascend, T-Head, Moore Threads, Cambricon, and others). Notable demos include porting the Apollo 11 guidance computer to Rust from scratch and an 880K-token continuous AutoClaw agent task. Remaining gaps: a 13% deficit to Opus 4.8 on SWE-Marathon, unknown 1M-context pricing, and unproven long-term stability on domestic hardware.

Event

On June 16, 2026, Zhipu AI launched and open-sourced GLM-5.2 under the MIT license. Key specs and results:

1M-token context window
51 points on the Artificial Analysis composite leaderboard — top three alongside Anthropic and OpenAI, open-source SOTA
Ranked #1 among all available models in Code Arena's million-user blind testing

Source page: https://www.zhipuai.cn/zh/research/161

Positioning: Long-Horizon Tasks

Zhipu devoted nearly all of 2025 to coding — from GLM-4.5 to 4.7 establishing itself as the strongest open-source coding model. With GLM-5.2, the goal shifts to enabling models to plan, execute, verify, and repair continuously over days or even weeks, like a top-tier engineer.

Core Breakthroughs

1. Solid 1M Context

1M context windows are common, but most degrade noticeably beyond a few hundred K tokens. Zhipu expanded its 1M coding-agent training environments from "large-repo completion" to real long-horizon scenarios such as large-scale implementation, automated research, and performance optimization. In practice, the 1M experience can sometimes rival Opus:

FrontierSWE: GLM-5.2 trails Opus 4.8 by only ~1 percentage point, ahead of GPT-5.5 and Opus 4.7
SWE-Marathon: still a 13% gap — the next challenge

2. Extreme Infrastructure

IndexShare: one attention indexer reused every four sparse-attention layers, compressing per-token FLOPs at 1M context to 2.9x
MTP speculative decoding: acceptance length improved by up to 20%
In-house Slime framework: supports large-scale agentic RL and OPD training

Day-0 Domestic Compute Support

GLM-5.2 ran on day one across nine domestic Chinese compute platforms: Huawei Ascend, T-Head (Pingtouge), Moore Threads, Cambricon, Kunlunxin, MetaX, Hygon, Biren, and Iluvatar CoreX. The Ascend 950 supernode launching in H2 is expected to become a strong backbone for GLM-5.2.

Engineering Highlights

Moonshot experiment: GLM-5.2 ported the Apollo 11 guidance computer (~4,600 lines of fixed-point CPU code) to Rust from scratch, running the original 65,000-line lunar landing program unchanged — autonomously reproducing the famous 1202 alarm that nearly aborted the landing.
AutoClaw: sustained an 880K-token continuous long-horizon task spanning web, mobile, and mini-program clients.

Why It Matters

1. First time a domestic Chinese coding model directly competes with Opus on long-horizon tasks — delivery capability, not benchmark gaming. 2. MIT license + Day-0 domestic compute: model, licensing, and silicon all ready simultaneously — a signal event for Chinese enterprise buyers. 3. The 1M context is engineered for real use, not just a spec-sheet number. 4. Adjustable effort levels let developers trade off capability, speed, and cost.

Risks and Open Questions

SWE-Marathon still lags Opus 4.8 significantly; ultra-long-cycle agents remain hard.
Day-0 support on domestic chips does not guarantee Day-30 stability — throughput, memory, and tail latency are the real industrial battleground.
The cost curve for 1M context is unknown; enterprise adoption depends on pricing.
Ecosystem overlap between AutoClaw/ZCode and Cursor/Claude Code — developers will vote with their feet.