Z.AI (ZHIPU AI) — Weekly Intelligence Report
Report Period: March 1–7, 2026 | Published: March 7, 2026 Coverage: HKEX: 2513 (Knowledge Atlas Technology)
At a Glance
| Metric | Detail |
|---|---|
| Flagship Model | GLM-5 (744B params / 40B active, MoE) |
| Model Release | February 11, 2026 |
| Stock Reaction | +28.7 – 34% (HKEX) on launch day |
| API Pricing | ~$0.80/M input · $2.56/M output tokens |
| Open-Source License | MIT — commercial use permitted |
| Compute Infra | Huawei Ascend (full US-hardware independence) |
| Context Window | 200K tokens (via DeepSeek Sparse Attention) |
| Key Incident | Late-Feb compute shortage; new-user signups restricted |
Executive Summary
Z.ai (HKEX: 2513, formally Knowledge Atlas Technology, formerly Zhipu AI) has spent the past month executing the most consequential moves in its short public-company history. The launch of GLM-5 on February 11, 2026 — the most capable open-source large language model benchmarked to date — set off a 34% single-session rally on the Hong Kong Stock Exchange and signalled that China’s AI labs have closed what analysts once estimated as a seven-month lag behind Western rivals to roughly three months.
In the reporting week of March 1–7, 2026, the post-launch operating reality is coming into sharper focus: aggressive user adoption drove the company to restrict new sign-ups and issue a public call for compute support in late February, while a 30% GLM Coding Plan price increase — China’s first major LLM tariff hike of 2026 — signals a historic pivot from user-acquisition subsidies to monetisation.
Three strategic dynamics dominate the outlook:
- China’s self-reliant AI-chip stack is proven at frontier scale.
- Open-source MIT licensing is being weaponised for global enterprise adoption.
- Z.ai is transitioning from benchmark competitor to agentic enterprise software provider.
⚡ Why It Matters: GLM-5 is the first open-weight model to achieve parity with Claude Opus 4.5 and GPT-5.2 on agentic engineering benchmarks — at ~6× lower API cost. It runs entirely on domestic Chinese hardware. For enterprises, investors, and policymakers, this changes the competitive calculus of the global AI race.
In-Depth Analysis
1. Strategic Context
Z.ai’s trajectory over the past 60 days represents the fastest maturation arc of any Chinese AI lab post-IPO. Having raised HKD 4.35 billion (~USD 558M) in its January 8, 2026 Hong Kong listing — the first Chinese LLM company to do so — the company deployed capital at speed: GLM-5 launched just 34 days later.
Three strategic priorities are now clearly legible:
- Frontier parity through architectural efficiency. The MoE design (40B active / 744B total) and the novel ‘Slime’ asynchronous RL training infrastructure replace brute-force compute with targeted post-training iteration, allowing Z.ai to compete with far better-resourced Western labs.
- Geopolitical insulation via domestic hardware. GLM-5 was trained entirely on Huawei Ascend chips — the first frontier-class model to achieve this at scale. In a world of tightening US export controls on GPUs, this represents a structural competitive moat no Western open-source competitor can replicate.
- Monetisation inflection. The 30–60% subscription hike and 67–100% API fee increase mark Z.ai’s deliberate exit from China’s 2024–2025 LLM price-war era. The company is betting that GLM-5’s technical differentiation can sustain premium pricing while remaining materially cheaper than Anthropic or OpenAI.
2. Market Impact
Stock & Investor Sentiment
The GLM-5 launch triggered a 34% single-session surge on the HKEX, dragging infrastructure partner UCloud Tech up 20% and SenseTime up 6.8% in sympathy. JPMorgan Chase issued a buy recommendation on Z.ai and MiniMax in February, reflecting renewed institutional confidence in Chinese AI equities after DeepSeek’s 2025 disruption narrative.
Compute Crisis as Demand Signal
In late February, Z.ai’s shares fell ~23% amid reports of compute resource shortages that forced the company to restrict new-user sign-ups. The dual signal — a selloff on supply constraints paired with evidence of overwhelming demand — mirrors the dynamic faced by Anthropic and OpenAI at earlier inflection points. For investors, the capacity crunch is a validation of product-market fit, not a fundamental risk.
Pricing Power Test
China’s AI market has been characterised by suicidal price competition since DeepSeek’s 2025 open-weight release. Z.ai’s 30% Coding Plan increase — framed around GLM-5’s compute costs — is the first credible test of whether any Chinese lab can charge premium prices. Early evidence suggests the market is absorbing the hike.
3. Technical Analysis — GLM-5 Architecture
| Parameter | Detail |
|---|---|
| Parameter Scale | 744B total / 40B active (MoE, 256 experts, 8 activated per token) |
| Training Data | 28.5T tokens pre-training corpus |
| RL Infrastructure | ‘Slime’ — novel asynchronous RL; 2× throughput vs. traditional approaches |
| Attention Mechanism | DeepSeek Sparse Attention (DSA); preserves 200K context, cuts deployment cost |
| Hallucination Rate | −1 on AA-Omniscience Index — 35-pt improvement; industry-leading abstention |
| SWE-bench Verified | 77.8% (#1 open-source; Claude Opus 4.5 at 80.9%) |
| AIME 2026 I | 92.7 (matches Claude Opus 4.5 at 93.3) |
| Humanity’s Last Exam | 50.4 (tool-augmented) — beats Claude Opus 4.5 at 43.4 |
| Hardware | Huawei Ascend / Moore Threads / Cambricon / Kunlunxin / MetaX + Nvidia |
The ‘Slime’ RL system is the most important under-reported innovation. By decoupling trajectory generation from parameter updates, Zhipu AI can run fine-grained post-training iterations at scale — the same RL-driven capability refinement that gave OpenAI’s o-series and Anthropic’s extended thinking models their edge. GLM-5 demonstrates this technique is now replicable by a well-funded Chinese lab with domestic hardware.
⚠️ Safety Note: Lukas Petersson (Andon Labs) noted that GLM-5 achieves goals through ‘aggressive tactics’ with limited situational awareness — a profile associated with instrumental convergence risks in long-horizon agentic deployments. Enterprises building critical workflows on GLM-5 should implement human oversight checkpoints.
4. Product Launch — Z.ai Agent Mode & Document Generation
Concurrent with GLM-5’s model release, Z.ai formally launched an enterprise-grade Agent Mode on its consumer and API platforms — a material product expansion, not merely a model upgrade:
- Native document generation — GLM-5 autonomously produces formatted .docx, .pdf, and .xlsx files from raw prompts or source material, including financial reports, PRDs, and sponsorship proposals.
- OpenClaw integration — GLM-5 is compatible with the cross-app/cross-device workflow framework, plus established coding agents including Claude Code, OpenCode, and Roo Code.
- Z.ai API platform expansion — live on OpenRouter and at api.z.ai, with Python and Java SDK support enabling enterprise developers to integrate GLM-5 into existing stacks.
- GLM Coding Plan — positioned explicitly as China’s answer to Anthropic’s Claude Code (unavailable in China). The 30% price hike on this plan signals Z.ai believes it has pricing leverage.
The Agent Mode launch frames Z.ai’s ambition clearly: the company is not positioning as a model provider to AI builders — it is positioning as an agentic enterprise software platform. This mirrors the strategic shift Anthropic made with Claude Code and OpenAI with Operator, and represents a fundamentally larger addressable market than API token sales.
Forward Outlook
The next 60 days will be critical for Z.ai on three fronts:
- Compute capacity resolution — the ability to lift sign-up restrictions and stabilise inference latency will determine whether the late-February selloff was a temporary dislocation or a signal of structural infrastructure risk.
- Enterprise adoption conversion — Agent Mode adoption metrics will determine whether Z.ai can justify its premium pricing tier. Early indicators from the GLM Coding Plan are positive; B-end (enterprise API) conversion rate is the key watch item.
- Western market traction — GLM-5’s MIT license and OpenRouter availability lower the barrier for US/EU enterprise trials substantially. Given the model’s cost-performance ratio (~6× cheaper than Claude Opus 4.6), expect significant developer experimentation in Q2 2026.
🔭 Analyst View: Z.ai has de-risked the ‘can Chinese labs match Western frontier models on domestic hardware’ question. The remaining questions are commercial: can they convert benchmark credibility into durable enterprise revenue at scale? The 2026 monetisation cycle will answer this.
Sources
| # | Source | Link |
|---|---|---|
| 1 | VentureBeat — GLM-5 launch & Slime RL | https://venturebeat.com/technology/z-ais-open-source-glm-5-achieves-record-low-hallucination-rate-and-leverages |
| 2 | SCMP — GLM-5 challenge to rivals | https://www.scmp.com/tech/article/3343239/chinas-zhipu-ai-launches-new-major-model-glm-5-challenge-its-rivals |
| 3 | CNBC — Chinese AI stock rally | https://www.cnbc.com/2026/02/12/chinese-ai-stocks-new-model-and-agent-releases-zhipu-minimax.html |
| 4 | Bloomberg — Lunar New Year AI race | https://www.bloomberg.com/news/articles/2026-02-11/china-s-zhipu-unveils-new-ai-model-jolting-race-with-deepseek |
| 5 | The Decoder — GLM-5 MIT release | https://the-decoder.com/chinese-ai-lab-zhipu-releases-glm-5-under-mit-license-claims-parity-with-top-western-models/ |
| 6 | Techloy — GLM-5 launch + 30% price hike | https://www.techloy.com/chinas-zhipu-ai-launches-glm-5-with-30-price-increase-as-stock-jumps-34/ |
| 7 | The Register — GLM-Image / Huawei hardware | https://www.theregister.com/2026/01/15/zhipu_glm_image_huawei_hardware/ |
| 8 | Wikipedia — Z.ai (compute shortage / share decline) | https://en.wikipedia.org/wiki/Z.ai |
| 9 | TrendForce — DeepSeek/Zhipu market context | https://www.trendforce.com/news/2026/02/13/news-deepseek-expands-context-tenfold-as-zhipu-rolls-out-new-model-in-chinas-ai-race/ |
| 10 | Build Fast With AI — GLM-5 benchmark deep-dive | https://www.buildfastwithai.com/blogs/glm-5-released-open-source-model-2026 |
