News
vLLM 0.28.0 ships Model Runner V2 defaults and breaks bitsandbytes-in-tree; Claude Code latest is 2.1.247 while stable stays 2.1.231.
vLLM v0.28.0 (26 Aug) raises max_num_batched_tokens from 8192 to 16384, enables Mamba prefix caching by default, and matures Model Runner V2 (E/P/D disaggregation, tiered KV offload including disk, multi-layer MTP KV).1 Breaking: bitsandbytes moves out-of-tree; Transformers floor is 5.15.0; calculate_kv_scales and override_attention_dtype are removed.1
npm @anthropic-ai/claude-code: stable still 2.1.231; latest/next → 2.1.247.23 2.1.247 adds SendFeedback drafts and fixes /compact / --agent summarization under the wrong system prompt.3 Channel policy from #015 stands.
Breaking
vLLM: checklist bnb plugin + Transformers ≥5.15 before upgrade. Claude Code: latest ≥2.1.247 for those fixes; stable remains 2.1.231.
Sources
- github — vLLM v0.28.0 · read
- npm — @anthropic-ai/claude-code dist-tags · read
- github — claude-code v2.1.247 · read
Originally part of the 27 August 2026 edition.