Show HN: Metis – An agent harness pushing DeepSeek to Opus-tier coding (82%)

3 pointsposted 12 hours ago
by oliverhuchenrui

3 Comments

oliverhuchenrui

10 hours ago

In the final benchmark tests, DeepSeek V4 Flash achieved a score of 82.02%, coming close to the 83.8% score of Claude Fable 5 XHigh—effectively giving a "Flash" model the capabilities of a flagship model.

ozozozd

9 hours ago

I think everyone will look for a comparison with Pi.