Jifu GLM-5: China's counterattack that created the world's third largest AI model with only Huawei chips

Updated: 2026-02-21 | Category: aiNews

1) Problem definition

• Target audience: Technology/business leaders, strategic planning officers, product/operations managers
• Solution Problem: Chinese AI startup Jifu has released GLM-5, a 745 billion parameter model learned only with Huawei Ascend chip. Even under US semiconductor regulations, it ranks third in the world in performance after Claude Opus 4.5 and GPT-5.2, changing the landscape of the competition for AI hegemony. We reconstruct this into actual decision-making and actionable standards.
• Scope: 2026-02-12 Convert to execution frame while maintaining the argument and context of the published article
• Exclusion range: unconfirmable rumors, exaggerated conclusions based on a single indicator, automated recommendations without verification

2) Evidence/Comparison (3 alternatives)

Alternative	Cost	Time	Accuracy	Difficulty	Recommended Situation
A. Keep the same way	Low~Medium	Start immediately	Low to medium (large deviation)	Low	When minimizing risk is a priority
B. Limited Pilot + Human Approval	Medium	2~6 weeks	Medium~High	Medium	The default choice for most organizations
C. Full introduction	High	1~3 months	High possible (governance premise)	High	Organizations with a mature standardization and audit system

• Judgment criteria: Cost (introduction + operation), time (lead time to realize value), accuracy (error rate/rework rate), difficulty (organizational change management)

3) Step-by-step execution (practical procedure)

1. Define goals: Numerically determine 1-2 current bottlenecks (time, quality, approval delays).
2. Data/evidence organization: Figures and cases used in existing articles are separated by source and verification status is displayed.
3. Pilot design: Assign one team of tasks (or one service) and fix the scope of the experiment for 2-4 weeks.
4. Execution Gate: Documents approval rules (reliability threshold, exception routing, rollback condition) before automatic processing.
5. Measures: Weekly tracking of at least 3 of the following: processing time, error rate, rework rate, and user satisfaction (CSAT/NPS).
6. Expansion/discontinuation decision: If KPI is met, expand; if not met, disassemble the cause (data/process/permissions) and re-experiment.

Execution example (common):

#1) Save pilot baseline
echo "baseline: lead_time,error_rate,rework_rate" > pilot-metrics.csv
#2) Cumulative weekly results
echo "week1,12h,2.4%,18%" >> pilot-metrics.csv

4) Pitfalls/Mistakes and Prevention/Recovery

1. Tool-centric introduction: If you introduce tools first without defining the problem, the ROI will be unclear.

• Prevention: Create decision-making documents in the order of problems-indicators-tools.

1. Automation without verification: Automated execution without confidence thresholds and approval mechanisms leads to quality incidents.

• Prevention: High-risk items force human approval (HITL).

1. No logs preserved: Results may look good, but no audit trail prevents operations from scaling.

• Recovery: Recollect input/output/approval history into standard log schema.

1. Exaggerated performance promotion: generalizing from short-term sample figures reduces credibility.

• Prevention: Sample number, period, and exclusion conditions are also disclosed.

5) Execution checklist (including DoD)

• Documented one target task and exclusion scope.
• Two or more alternatives were compared in terms of cost/time/accuracy/difficulty.
• Defined authorization rules (reliability threshold, exception routing, rollback).
• Track 3 or more KPIs (time/error/rework/satisfaction) weekly.
• There is a prevention/recovery runbook for 3 or more failure patterns.
• Reference material link and confirmation date are specified in the text.
• Author recommended/not recommended/conditional exception recorded.

**Definition of Done:** Improved at least 2 key KPIs in a 2+ week pilot + 0 quality/security incidents + Approved by Operations Director

6) References (link + date)

• Reuters AI News Hub: https://www.reuters.com/technology/artificial-intelligence/ (Confirmation date: 2026-02-21)
• OECD AI Policy Observatory: https://oecd.ai/ (Confirmation date: 2026-02-21)
• NIST AI RMF 1.0: https://www.nist.gov/itl/ai-risk-management-framework (Confirmation date: 2026-02-21)
• UN AI Advisory Body data: https://www.un.org/en/ai-advisory-body (Confirmation date: 2026-02-21)

7) Author's perspective

• Recommendation: Introduce steps based on pilot metrics and operational logs rather than exaggerated single numbers.
• Non-recommendation: This is a method of deciding on introduction/discontinuation based solely on unsourced claims or provocative headlines.
• Conditional exception: Organizations with high regulatory demands and already mature audit systems can expand the scope of automation more quickly.

---

Summary of existing issues (preservation)

Chinese AI startup Zifu unveiled GLM-5, a 745 billion parameter model learned only with Huawei Ascend chip. Despite U.S. semiconductor regulations, it is changing the landscape of the competition for AI hegemony by ranking third in the world after Claude Opus 4.5 and GPT-5.2.