The skyrocketing AI project J-Space has been accused of fabricating data, as the community's retesting results are "completely opposite" to the official data.

BlockBeats News, August 18th. According to MaxForAI, today, the AI project J-Space Cognition Suite, which recently gained popularity on the X platform, faced community scrutiny. It was alleged that the claimed test results of DeepSeek V4 by the project could not be reproduced. Previously, the project stated that V4 Flash, when combined with J-Space, could match GLM-5.3, while V4 Pro could outperform Fable 5 on multiple Agent Benchmarks, with a 2.53x speed increase and a 2.21x Token efficiency improvement.

GitHub user GoForceX conducted a high-concurrency retest using 87 questions from Terminal Bench 2.1 and confirmed the integration of J-Space-related modules. The results showed that after incorporating J-Space, the benchmark results slightly decreased, and both Token usage and costs increased, contradicting the project's claims of performance, speed, and Token efficiency improvement.

Subsequently, the community requested the project to publicly disclose complete evaluation configurations, per-question results, execution logs, original runtime, and Token consumption data. Currently, the project has mainly released summary results and has not provided comprehensive raw experimental records sufficient to validate the precise data mentioned above. It is worth noting that in response to the criticism, the project's authors admitted that the related data was "indeed exaggerated" and indicated that the actual improvement was roughly between 1.6x and 3x. As of now, the authors have not issued a formal response to the community's doubts, and some related questioning issues have been deleted.