The Circuitry
THE CIRCUITRYYour one-stop source for all tech news
HOMETODAYNEWSFEEDEVENTS
BOOKMARKS
RSS
© 2026 The Circuitry
About UsSourcesContactCorrectionsPrivacy
  • Today
  • Feed
  • Events
  • Saved
Scroll for more
Verification
VERIFIEDConfidence: HIGH
Source identified
Claims cross-referenced
No discrepancies found
Fact-check summary

Z.ai's GLM-5.2 release, benchmarks, Huawei training, pricing, and 1M context are corroborated by VentureBeat, the company's official blog, Hugging Face, docs.z.ai, Wikipedia, and multiple tech outlets as of June 16-18 2026.

Sourcing
1source

via Decrypt

Decrypt · track record
31Stories
100%Verified
230d
All sources →
Home/Tech/Z.ai Launches GLM-5.2, a Huawei-Trained Rival to Top AI Models
VERIFIEDBy Xavier Rivera· ·2 min read

Z.ai Launches GLM-5.2, a Huawei-Trained Rival to Top AI Models

Z.ai released GLM-5.2 on June 16, which scores within 1 percent of Claude Opus 4.8 on FrontierSWE while beating GPT-5.5. The MIT-licensed model trained solely on Huawei Ascend chips without NVIDIA hardware and undercuts Western API pricing.

Source:Decrypt
Post
Z.ai Launches GLM-5.2, a Huawei-Trained Rival to Top AI Models
TL;DRAI · 60 sec read

Z.ai releases GLM-5.2, a 744B MoE model trained solely on Huawei Ascend chips. It scores 74.4 on FrontierSWE, within 1 percent of Claude Opus 4.8 and above GPT-5.5, while offering a 1M-token context. The open-source model undercuts Western API prices and faces no regional limits, though local runs need 256GB memory.

Z.ai has introduced GLM-5.2, which reportedly delivers performance within 1 percent of Claude Opus 4.8 on the FrontierSWE benchmark while surpassing GPT-5.5.

GLM-5.2 posts strong benchmark scores. The model achieved 74.4 on FrontierSWE against Claude Opus 4.8's 75.1 and GPT-5.5's 72.6. It reached 62.1 on SWE-bench Pro, exceeding both GPT-5.5 at 58.6 and its predecessor GLM-5.1 at 58.4.
Development occurred solely on Huawei Ascend chips without any NVIDIA components.
These results position it as the leading open-source entry on the Artificial Analysis Intelligence Index that combines nine separate evaluations. Benchmarks from OpenRouter align it with the now-banned Claude Fable 5.

The model was trained exclusively on Huawei silicon. Development occurred solely on Huawei Ascend chips without any NVIDIA components. Stability AI founder Emad Mostaque placed overall training expenses near $25 million, approximately 80 percent of which went toward post-training.
From The CircuitryThe Feed — live briefs across tech, all day.See what’s happening →
The Beijing laboratory, listed on the U.S. Entity List since January 2025, had already used Huawei Ascend Atlas servers for image models without American hardware. GLM-5.2 expands that foundation as a 744-billion-parameter mixture-of-experts system featuring a genuine 1-million-token context window, five times the 200K capacity of GLM-5.1.
GLM-5.2 expands that foundation as a 744-billion-parameter mixture-of-experts system featuring a genuine 1-million-token context window, five times the 200K capacity of GLM-5.1.
Pricing undercuts Western frontier models. Access through the API runs $1.40 per million input tokens and $4.40 per million output tokens. Those figures stand well below Claude Opus 4.8 rates of $5 input and $25 output. A monthly Coding Plan begins near $18 and works inside Claude Code, Cline, Kilo Code plus leading agentic platforms. The release carries an MIT license and carries no regional limits.

Local deployment requires substantial hardware. Unsloth AI supplied 2-bit GGUF quantizations that compress the original 1.51TB file size down to 238GB while preserving roughly 82 percent accuracy. Execution still calls for 256GB of unified memory or an equivalent RAM/VRAM setup, whether a fully equipped M4 Ultra Mac Studio or a mid-range GPU workstation paired with 256GB system RAM and mixture-of-experts offloading.
Z.ai released GLM-5.2 on June 16. Combined with the recent ban on Anthropic Fable, the debut has reportedly lifted the company's shares 90 percent to a record high.
Why this mattersAI · ~100 words

Tap a lens to see what this story means for you.

Reader-supported
DonateBuy me a coffee →Follow@thecircuitry_ →Follow@thecircuitry.to →

Reader-supported · The Brief

Liked this? The Brief brings you the whole day in tech, verified, every morning. Two minutes, free forever.

HELP US IMPROVE
From The Circuitry

See what’s happening right now

The Feed runs all day — short, verified briefs the moment they break.

Open the Feed →
From The Circuitry

Follow @thecircuitry_

Every story we publish, as it happens. No noise between.

Follow on X ↗On Bluesky ↗

Reader-supported

The Circuitry is a passion project I've always wanted to build, and I love the work behind it.

Running it costs real money. APIs, hosting, time. To keep improving the site and growing this into something useful for everyone, those costs have to be covered.

Any contribution is appreciated. If not, no pressure. Thanks for reading.

Buy me a coffee
AIOpen SourceChina
More fromDecrypt
  • Robinhood Chain Launches as Ethereum L2 for Tokenized Stocks

    Markets · 27d
  • Robinhood Chain racks up $1B DEX volume in debut week

    Markets · 29d
  • Paradigm Closes $1.2 Billion Fund Aimed at Crypto, AI and Robotics Startups

    Markets · 1mo
More inTech
  • Tesla and SpaceX confirm Terafab chip fab in Texas

    Tech · 1d
  • OpenAI Urges Federal Judge to Throw Out Apple's Trade Secrets Complaint

    Tech · 2d
  • Meta introduces Muse Code, its terminal-based coding agent

    Tech · 2d
SupportThe Work

The Circuitry is reader-supported. If you find the daily brief useful, you can buy me a coffee to keep it going.

Buy a coffee →
SubscribeCircuitry Brief

Liked this? The Brief brings you the whole day in tech, verified, every morning. Free forever.

MORE IN TECH

Tesla and SpaceX confirm Terafab chip fab in Texas

Tesla and SpaceX have confirmed Grimes County, Texas as the site for their Terafab semiconductor megafactory, with the first phase costing roughly $16.8 billion. The project targets the largest chip manufacturing facility on the planet to supply over 1 terawatt of compute per year that exceeds current and future global production capacity.

OpenAI Urges Federal Judge to Throw Out Apple's Trade Secrets Complaint

OpenAI has filed a motion asking a federal judge to dismiss Apple's trade secrets lawsuit, describing the claims as meritless. The dispute, which follows a July suit and this week's injunction request from Apple, highlights tensions after their prior partnership on Siri and OpenAI's hardware push.

Meta introduces Muse Code, its terminal-based coding agent

Meta has released an early beta of Muse Code, a terminal-based coding agent driven by the updated Muse Spark 1.2 model and positioned against Anthropic's Claude Code and OpenAI's Codex. Substantially lower rates, including a contributor plan at $0.10 for every million tokens received, may encourage migration away from higher-priced options such as Anthropic's Sonnet 5.