Breaking

Zhipu AI Releases Open-Weight GLM-5.2 With 1M-Token Context

July 6, 2026 at 17:45 EDT

Foundation Models
Open Source
Software Dev & Coding

China's Zhipu AI (Z.ai) released GLM-5.2, an open-weight large language model capable of handling a 1-million-token context window, in mid-June 2026. Distributed with fully open weights under an MIT license, the model is being promoted for its free local execution and its "long-horizon" performance in carrying out complex, extended tasks. According to the official announcement, the model is distributed via Hugging Face and other platforms.

June 30, 2026 · Meituan

LongCat-2.0: a 1.6-trillion-parameter open model, trained without a single NVIDIA GPU

China's food-delivery giant releases a near-frontier agentic coding model under an MIT license — with a 1-million-token context window and training completed entirely on domestic Chinese chips.

1.6T

Total parameters (MoE)

Token context (native)

30T+

Training tokens

~50K

Domestic chips, 0 NVIDIA GPUs

Agentic coding, near the frontier

SWE-Bench Pro — real-world code-fixing. LongCat-2.0 edges close to GPT-5.5-class results.

59.5

LongCat-2.0

58.6

GPT-5.5 class

Terminal-Bench: 70.8 ~48B active params / token MIT license — commercial OK

WHY PRACTITIONERS ARE EXCITED

Feed an entire website, competitor data & SOPs at once
"Remembers the whole project" across long documents
Open-source, free & local/offline potential for privacy

WHERE CAUTION REMAINS

Full weights not yet out — local use still limited
Trails GLM 5.2 on some tasks (e.g. game generation)
API geo-restrictions & domestic-chip setup hurdles

Part of a wave of powerful, open-weight Chinese models — DeepSeek, Qwen, GLM, Kimi — now spreading even to Silicon Valley startups.

Anonymously launched as "Owl Alpha," it topped OpenRouter leaderboards for ~2 months before being revealed as Meituan's work. The focus now shifts to real-world deployment.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.