实测星火X2.5：手搓粒子月亮、拆完61页财报……还揪出了我的Bug
Ranking
Overall
64
Content
70
Popularity
N/A
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR - iFlytek launched Spark X2.5, a 293B-parameter MoE model focused on coding and agentic task completion. Hands-on tests suggest it can turn complex, multimodal instructions into deliverables such as interactive websites, financial analyses, research drafts, and visualized reports.
- Spark X2.5 uses a 293B-A30B MoE architecture, supports more than 200 languages, and exposes Anthropic- and Responses-compatible APIs for agent frameworks including Claude Code and Codex.
- Tests covered generating gesture-controlled 3D webpages and a game website, analyzing a 61-page Nvidia earnings report, synthesizing research materials, and processing e-commerce data embedded as images in Word.
- The model reportedly detected fabricated survey data limitations and deliberately incorrect revenue calculations rather than simply accepting the supplied inputs.
- iFlytek says the model’s training, reinforcement learning, iteration, and inference run on domestic Chinese computing infrastructure; it previously reported raising comparable MoE training efficiency from 30% to 93% of an A800 cluster.
Sources (1)
实测星火X2.5：手搓粒子月亮、拆完61页财报……还揪出了我的Bug
Public signals
N/A
TL;DR - iFlytek launched Spark X2.5, a 293B-parameter MoE model focused on coding and agentic task completion. Hands-on tests suggest it can turn complex, multimodal instructions into deliverables such as interactive websites, financial analyses, research drafts, and visualized reports.
- Spark X2.5 uses a 293B-A30B MoE architecture, supports more than 200 languages, and exposes Anthropic- and Responses-compatible APIs for agent frameworks including Claude Code and Codex.
- Tests covered generating gesture-controlled 3D webpages and a game website, analyzing a 61-page Nvidia earnings report, synthesizing research materials, and processing e-commerce data embedded as images in Word.
- The model reportedly detected fabricated survey data limitations and deliberately incorrect revenue calculations rather than simply accepting the supplied inputs.
- iFlytek says the model’s training, reinforcement learning, iteration, and inference run on domestic Chinese computing infrastructure; it previously reported raising comparable MoE training efficiency from 30% to 93% of an A800 cluster.