ds greatly exceeds expectations, v4. Flash official version capability approaches opus 4.8, optimistic about accelerated growth of domestic models in 26H2 [Tianfeng Computer Miao Xinjun Team]

English summary

# Capability: Benchmark tests far exceed V4-Pro-Preview. DeepSeek V4. Flash official version, as a 284b small model, ranks among the first tier of domestic open-source agentic models. In agent programming tests, it leads GLM-5.2 but is weaker than k3; looking overseas, Agent Coding capability is close to GPT-5.6. Luna, overall capability slightly below Opus 4.8 and Luna, further below Terra and Sol. # Price: Inference price is only one-third of V4. Pro, 1/3 of GPT-5.6. Luna (after price reduction), one-tenth of Terra, and ~1/100 of Sol and Opus, continuing the extreme cost-effectiveness advantage of domestic models. # Implementation path: V4. Flash and V4. Pro belong to the same generation, same technical route, with little difference in architecture paradigm...