Index Company
Zhipu
17 CÈ pieces since 9 May 2026: 15 intel items, 2 essays. Newest first. All companies →
-
MIT weights land half a point behind Claude Opus: GLM-5.3-Flash costs $0.15 per million tokens
Z.ai released a 320B mixture-of-experts activating 18B parameters with a 1,048,576-token context under an MIT licence. It scores 84.3 on Terminal-Bench 2.1 against Claude Opus 4.8's 85.0, and lifts DeepSWE 37% over GLM-5.2.
-
A 753B model on one desktop GPU: FreeToken trades VRAM for 512GB of host memory
UC Berkeley and UT Austin researchers released FreeToken under Apache-2.0, a mixture-of-experts serving engine that splits work between GPU and CPU. It runs 753B GLM-5.2 on one workstation GPU and 35B at 39.3 tokens per second on an 8GB laptop.
-
Some of these bugs hid since 1981: Zhipu's GLM-5.3 surfaces 2,436 vulnerabilities
Working with Tsinghua, Nankai, and six Chinese security firms, Zhipu's GLM-5.3 surfaced 2,436 confirmed vulnerabilities across 269 projects, 1,097 of them medium-to-high severity. The flaws had escaped human review for 26.6 years on average.
-
Better agents without a bigger base: Z.ai's GLM-5.3 lifts Terminal-Bench from 4.6 to 28.3
Z.ai shipped GLM-5.3 on the same 743B base model as GLM-5.2. Every gain came from scaled post-training across more task environments; DeepSWE rose from 46.2 to 66.9, and the weights follow in about two weeks.
-
When open weights stop crossing borders, who loses? Beijing weighs curbs with Alibaba and Zhipu
China's commerce ministry has consulted Alibaba, ByteDance, and Zhipu AI on restricting overseas downloads of advanced model weights and transfers of training data, leaving cloud and API access to the same models open. Nothing is final.
-
Beijing Will Not Wall Off Its AI (Even Now That It Leads)
Days after an open-weight Chinese model took the top of the coding charts, Beijing began consulting its champions on export-controlling AI, the same weapon it called economic bullying when Washington aimed it the other way.
-
Beijing weighs walling off the open weights that seeded Asia: curbs floated for Qwen, Doubao, and GLM-5.2
Reuters reported on July 7 that Chinese ministries met Alibaba, ByteDance, and Z.ai about limiting overseas access to advanced AI models, including open-weight releases. Officials floated bars on public release and treating model theft as a national security crime. Nothing is decided.
-
A 295B open model that beats GPT-5.5 on science: Tencent ships Hy3 under Apache 2.0
Tencent open-sourced Hy3, a 295B-parameter mixture-of-experts model activating 21B per token, under Apache 2.0 with no geographic carve-out. It processes documents with 47% fewer tokens than the far larger GLM-5.2 and tops GPT-5.5 on a scientific-research benchmark.
-
Chinese open models from DeepSeek and Z.ai reach US frontier parity
US enterprises are moving workloads onto open-weight models from DeepSeek and Z.ai, now rated competitive with leading American systems, as OpenAI and Anthropic inference prices keep climbing.
-
Zhipu releases GLM-5.2 open weights, topping GPT-5.5 on SWE-bench Pro
GLM-5.2 ships under an MIT license with a 744-billion-parameter mixture-of-experts design, roughly 40 billion active per token, and a one-million-token context window. It scores 62.1 on SWE-bench Pro, above GPT-5.5's 58.6.
-
Chinese open-weight models pass 60 percent of OpenRouter's token volume
By May 2026, openly licensed models from Alibaba, DeepSeek, Moonshot, and Zhipu accounted for roughly 61 percent of tokens consumed on OpenRouter, with four of the five most-used models built in China.
-
Zhipu releases GLM-5.2 open weights, topping openly available model rankings
Zhipu released GLM-5.2 under an unrestricted MIT license, ranking as the top openly available model on Artificial Analysis' Intelligence Index while undercutting Claude Opus by roughly five times on price per token.
-
Coinbase halves its AI bill by defaulting engineers to Chinese open models
Coinbase cut internal AI spending nearly 50 percent by defaulting engineers to open-weight GLM-5.2 and Kimi K2.7 Code, priced near $1.40 per million tokens against Anthropic Opus at $5.
-
The best AI model in the world is now off-limits to everyone in Southeast Asia
On 12 June Washington ordered Claude Fable 5 disabled for every foreign national; the outage was narrow, but the precedent is a supply-chain lesson every operator already knows.
-
Zhipu AI open-sources GLM-5.2, trailing Claude Opus by one point on coding
Released June 13 under an MIT license, the 744-billion-parameter mixture-of-experts model carries a one-million-token context window and lands within one percentage point of Claude Opus 4.8 on FrontierSWE, the hours-long coding benchmark.
-
Four Chinese labs ship frontier-parity coding models in three-week window
Alibaba's Qwen 3.7 Max, Moonshot's Kimi K2.7, MiniMax M3, and Zhipu's GLM-5.2 all launched in May and June 2026, matching Claude Opus 4.8 and GPT-5.5 on coding benchmarks at one-fifth to one-thirtieth the per-token cost.
-
Four Chinese labs ship frontier-grade open-weights coding models in 12-day window
Z.ai's GLM-5.1, MiniMax M2.7, Moonshot's Kimi K2.6 and DeepSeek V4 all landed within a fortnight, hitting near-frontier benchmarks on agentic coding at under a third of Claude Opus 4.7's inference cost.