As The Geek Learns AI Engineering Digestfrom As The Geek Learns

Wednesday, August 26, 2026

vLLM 0.28.0 lands as Gemini Enterprise adds Legal and Financial Services previews

vLLM cut v0.28.0 this morning (5:46 AM ET) with a Kimi-K3 performance pass, DFlash2, Model Runner V2 E/P/D, and new model support. llama.cpp nightly b10632 adds Metal chunked SSD MMA for Mamba-2 prefill. Google Cloud posted Gemini Enterprise Legal and Financial Services packs yesterday after Tuesday’s close (preview). IBM Granite Speech 5.0 TurboCTC (470M English ASR) landed on Hugging Face Tuesday afternoon. Ollama v0.33.0 is now a stable tag. MLX itself is still v0.32.2.

Window ~2026-07-30 to 2026-08-26. Tuesday already covered llama.cpp 0.3.0 Metal FA, MLX 0.32.2, and AWS ARD. Primary sources only. Vendor evals flagged.

Breaking AI News

vLLM v0.28.0: Kimi-K3 pass, DFlash2, Model Runner V2

Published 2026-08-26T09:46:30Z (Wed 5:46 AM ET). 584 commits, 270 contributors. Kimi-K3: Decode Context Parallel, fused FlashKDA, combined all-gathers with a project-reported 1.5–3× kernel-level speedup, adaptive speculative budget with a project-reported ~60% better DSpark TTFT, optional shared-expert sharding saving a project-reported ~17 GiB per GPU. DeepSeek V4 sparse MLA end-to-end. New models: Muse Glimmer, Ling 3.0 Flash, Dots3 NOTE, Interns2mobius. Default max_num_batched_tokens 8192 → 16384. Breaking: bitsandbytes is out-of-tree; Transformers 5.15.0. Audio sample-rate DoS fix. pip install vllm. vendor

v0.28.0

Gemini Enterprise Legal and Financial Services (preview)

25 Aug, after Tuesday’s ~8:30 ET close. Two industry packs on the existing Gemini Enterprise control plane (VPC-SC, CMEK, private isolation, citation-backed grounding). Legal is a plugin in the Gemini Enterprise app; MCP connectors inherit matter-level permissions from iManage, NetDocuments, Everlaw, RelativityOne, Docusign, HighQ, Harvey, and others. Financial Services centers a Google-managed Financial Research agent (more than 50 skills, A2A, citations). Both preview. Workflow speed claims are vendor.

LegalFS

Enterprise

2 stories
2 stories Open section

Local

3 stories
3 stories Open section

MLX

1 story
1 story Open section

Try this on a Mac

1 story
1 story Open section

OSS tools

1 story

vLLM v0.28.0

Kimi-K3, DeepSeek V4 sparse MLA, DFlash2, Model Runner V2 E/P/D. Muse Glimmer and Ling 3.0 Flash land in the serving stack. Kernel / TTFT / memory numbers are vendor. vendor

1 story Open section

OSS models

1 story
1 story Open section