- Load
- Stream
- Execute
- Certify
- Full-weight checkpoint
- streams the 1.6T DeepSeek-V4-Flash-0731 FP4 weights
- Measured final median
- 13.126 seconds for one token across 43 layers, about 0.0762 tok/s
- Narrow certificate
- 43/43 layers match the captured fields for token ID 0 and one greedy step
01
localMoE
- Full-weight checkpoint streams the 1.6T DeepSeek-V4-Flash-0731 FP4 weights
- Bounded working set loads one layer and six routed experts at a time
- Rust-owned CUDA lifetime keeps streams, events, buffers, cache slots, and cuBLAS state request-scoped
- Measured final median 13.126 seconds for one token across 43 layers, about 0.0762 tok/s
- Narrow certificate 43/43 layers match the captured fields for token ID 0 and one greedy step
