xxr3376

7 ms·

1. 1.▲
   by xxr3376 8mo ago

   You can try Kimi's Anthropic-compatible API. Just connect Claude Code to Kimi's API endpoint and everything works well https://www.kimi.com/code/docs/en/more/third-party-agents.ht...
2. 2.▲
   K2VV: Wild Precision Gaps Across "Kimi K2" API Vendors(github.com)

   4 points by xxr3376 1y ago|1 comments
3. 3.▲
   by xxr3376 1y ago

   K2-Vendor-Verifier, a dead-simple harness that fires the same 2 000 tool-calling prompts at every vendor who claims to sell “Kimi K2” endpoints. The delta is brutal: the best node returns 99.3 % token-perfect answers, the worst barely hits
4. 4.▲
   by xxr3376 4y ago

   MegEngine team speedup depthwise Conv 10 times faster, which makes increasing kernel size almost cost-free while gaining better neural network accuracy.
5. 5.▲
   Speedup depthwise conv 10 times(medium.com)

   1 points by xxr3376 4y ago|1 comments
6. 6.▲
   by xxr3376 5y ago

   MegEngine (A Deep Learning Framework) implements DTR. Now you can train 3x larger model by tradeoff a little bit speed for lots of GPU memory.
7. 7.▲
   Training 3x larger model on the same GPU cards(github.com)

   2 points by xxr3376 5y ago|1 comments