vLLM Kimi-K3 and self-improving Claude agents #124
Today's Letter
vLLM v0.28.0, Kimi-K3 optimization release
- vLLM v0.28.0 includes 584 commits from 270 contributors, including 76 new contributors.
- Kimi-K3 adds DCP, fused FlashKDA kernels, SiTU, GEMM-RS, and combined all-gathers.
- Combined all-gathers deliver 1.5–3x kernel-level speedups.
- Adaptive speculative token budgets improve DSpark TTFT by about 60%.
- Shared-expert sharding can save about 17 GiB of memory per GPU.
- Kimi-K3 now runs on ROCm with the Model Runner V2.
- DeepSeek V4 supports sparse MLA for decode, MTP, and DSpark speculative decoding.
- Speculative decoding adds DFlash2 local convolution and candidate selection.
- Model Runner V2 adds E/P/D disaggregation, weight offloading, and multi-layer MTP KV cache support.
Source: github.com
Warp builds self-improving agents on Claude
- Warp uses Claude Platform to build agents that improve from accumulated user feedback.
- The approach addresses low-quality output from its internal code review agent.
- Feedback is encoded in file-based Agent Skills so it persists beyond individual sessions.
- Warp’s stack includes Rust, Golang, GitHub Actions, and its internal Oz orchestration platform.
- Warp reports $73M raised, 800K monthly developers, and 10M Claude Code sessions to date.
Source: claude.com
More: youtube.com
Jocoletter curates AI, software, and product trends for developers and builders.
#Warp #vLLM