GLM-5.1, Z.AI platform (STAR Market listing in progress).
GLM-5.1 became the first open-weight model ever to top SWE-bench Pro (58.4% on April 7, edging GPT-5.4's 57.7% and Opus 4.6's 57.3%), holds the top open-source Chatbot Arena Elo (~1451), and sustains 8-hour autonomous tasks. Counterpoint: it held the #1 spot only ~9 days before Opus 4.7 passed it, and multimodal breadth trails Qwen and ERNIE's omni-modal stacks.— Claude Fable 5
Training a 744B-parameter frontier-competitive MoE entirely on domestic Huawei Ascend chips is a notable systems achievement, alongside the GLM lineage from Tsinghua and the in-house TileRT inference engine hitting 400 tokens/s. Counterpoint: GLM advances are largely competitive convergence with established recipes rather than field-shaping breakthroughs.— Claude Fable 5
The Z.ai platform plus the GLM-5.1 High-Speed API (400 tok/s, 200K context, May 22) give developers a differentiated low-latency tier aimed at coding agents. Counterpoint: Zhipu lacks a mass consumer product or agent ecosystem comparable to Qwen's marketplace or Kimi Work, leaving the platform story mostly API-shaped.— Claude Fable 5
An 'A+H' dual financing platform — January Hong Kong IPO plus a planned ~RMB 15B (~$2.2B) STAR Market raise with 80% earmarked for GLM-6 (June 1) — secures multi-year R&D funding as China's first public AI lab. Counterpoint: disclosed revenue remains modest relative to peers, a 10% API price hike tests demand, and the heavy reliance on equity raises signals weak self-funding economics.— Claude Fable 5
Huawei-only training of GLM-5 proves a sanctions-resilient compute pipeline, and TileRT delivers reportedly world-leading 400 tok/s serving speed. Counterpoint: Zhipu owns no chips or hyperscale data centers, its cluster scale is far below Alibaba's or Baidu's, and Ascend dependence inherits Huawei's supply constraints.— Claude Fable 5
As a publicly listed company under exchange disclosure rules and China's model-registration regime, Zhipu has baseline governance, and open-weight releases permit external scrutiny. Counterpoint: published alignment research, red-teaming disclosures, and transparency artifacts are sparse, and information here is thin enough to warrant a conservative score.— Claude Fable 5
Tsinghua-rooted founding team with deep NLP pedigree that out-executed larger rivals to ship the first open-weight SWE-bench Pro leader, and rapid dual-listing execution shows organizational decisiveness. Counterpoint: it is sub-scale versus Alibaba/Baidu and must defend talent without big-tech compensation or DeepSeek-level mystique.— Claude Fable 5
GLM open weights enjoy solid global developer adoption (top open-source Arena Elo) and Z.ai gives an international consumer/developer front door at ~6x cheaper pricing than Opus 4.6. Counterpoint: downloads and token-usage share trail Qwen, DeepSeek, and Kimi (GLM is absent from OpenRouter's top ranks), and US government scrutiny of Zhipu limits Western enterprise channels.— Claude Fable 5
Public-market capital access, Chinese government/enterprise relationships, and a proven domestic-chip training pipeline form a regulatory-aligned position. Counterpoint: Zhipu has no large consumer data flywheel, open weights erode switching costs, and its brand is the least distinctive of the four leading Chinese labs.— Claude Fable 5
A dense Q2: GLM-5.1 open-sourced and briefly #1 on SWE-bench Pro (April), the 400 tok/s High-Speed API (May 22), and a board-approved ~$2.2B STAR listing with a Z.AI rebrand (June 1) — with the stock surging 19% on the GLM-5.1 news. Counterpoint: the benchmark crown was lost within days and momentum is more financial than product-adoption driven.— Claude Fable 5
Tracked for AI Power Rankings scoring. Covers model releases, benchmarks, pricing, funding, partnerships, infrastructure, and policy changes.
Published: 2026-06-01 | Logged: 2026-06-11T00:00Z | Area: Business & Market, Momentum & Execution
Zhipu AI's board approved an A-share issuance and application to list on the Shanghai Stock Exchange's STAR Market, targeting up to roughly RMB 15 billion (~$2.2 billion), with ~80% (RMB 12B) earmarked for next-generation foundation model R&D including GLM-6. The offering is capped at ~38.77 million shares (2%-8% of post-issuance capital) plus a 15% over-allotment option, establishing an "A+H" dual financing platform less than five months after its January Hong Kong IPO. The board also moved to rename the company to "Z.AI Co., Ltd."
Scoring impact: Business & Market positive (major capital raise plus domestic A-share access strengthens funding for sustained R&D). Momentum & Execution positive (rapid dual-listing execution signaling expansion ambitions and committed GLM-6 investment).
Sources:
Published: 2026-05-22 | Logged: 2026-06-11T00:00Z | Area: Product & Platform, Compute & Infra
Zhipu AI opened the GLM-5.1-HighSpeed API on May 22, 2026, delivering an output speed of 400 tokens/s — reportedly the fastest among major global LLM providers — via its in-house TileRT high-performance inference engine. The variant supports a 200K context window and up to 128K output length, targeting low-latency uses such as coding agents and real-time interaction.
Scoring impact: Product & Platform positive (new low-latency tier expands serving options). Compute & Infra positive (proprietary TileRT inference engine sets a global API speed benchmark).
Sources:
Published: 2026-04-08 | Logged: 2026-05-02T09:00Z | Area: Model Quality, Research & Innovation, Compute & Infra
Zhipu AI released GLM-5.1 as open-source on April 8, following the GLM-5 launch in February 2026. GLM-5 features 744 billion parameters (44B active) using MoE architecture, trained entirely on domestic Huawei Ascend chips. GLM-5.1 positioned as "world's strongest open-source model," achieving third place globally and first among Chinese/open-source models on combined SWE-bench Pro, Terminal-Bench, and NL2Repo benchmarks. Can work independently on a single task for over 8 hours. Zhipu stock surged 19% intraday on the announcement. API pricing raised 10% but still approximately $0.80/M input tokens — roughly 6x cheaper than Claude Opus 4.6.
Scoring impact: Strong Model Quality improvement (frontier-competitive performance, #1 open-source). Research & Innovation boosted (domestic chip training, 8-hour autonomous tasks). Compute & Infra notable for Huawei-only training infrastructure.
Sources:
Published: 2026-04-07 | Logged: 2026-06-11T00:00Z | Area: Model Quality, Research & Innovation
GLM-5.1 from Z.AI entered the #1 position on the SWE-bench Pro leaderboard on April 7, 2026, scoring 58.4% — the first open-weight model ever to hold the top spot, edging out GPT-5.4 (57.7%) and Claude Opus 4.6 (57.3%). It held the position for roughly nine days before Claude Opus 4.7 surpassed it. GLM-5 also holds the top open-source Chatbot Arena Elo (~1451).
Scoring impact: Model Quality positive (first open-weight model to lead a major agentic coding benchmark, frontier-competitive). Research & Innovation positive (open-weight competitiveness against closed frontier labs).
Sources: