Part 3 of assembling a DeepSeek-V4.1-Flash server on two RTX PRO 6000s. A host-expert plan built from our own agent traffic cut time per verify step by 2.42% in 5 of 5 randomized blocks and is served. A newer vLLM base took 10.0% longer per agent-task draw in 4 of 4 blocks and was not adopted. Our first randomized agent-task quality check of a served part was inconclusive at its planned size, and a time split of today's stack points our first own part at decode.

Part 2 of assembling a DeepSeek-V4.1-Flash server from the best parts of other engines on two RTX PRO 6000s. An SM120 mHC overlap from kacper-daftcode's vLLM-Moet cut time per verify step by 4.88% in 5 of 5 randomized blocks and is now served for single-stream use. The CPU-expert schedule passed its ten-block agent-task comparison with no pass-rate harm shown, the prompt region cache is served, and four things did not work.

We are assembling a server for DeepSeek-V4.1-Flash on one workstation with two RTX PRO 6000s from the best parts of other engines. Part 1: the body (vLLM with diffbot's EXL3 3-bit pack), one CPU-expert schedule served at 2.07% less time per verify step, and a prompt region cache that passed its test at 5.57 s less server time per agent-task draw and is next in line.

The released JetSpec++ drafter for Qwen3.8-27B beat our MTP head on short math prompts. On 77K–224K-token prompts with thinking on, it accepted 1.1–1.2 tokens per step and decoded 3.6–5.6x slower. An FP32 reference check found no large defect in our port on the prompt we checked.

Kernels, 4-bit quantization and speculative decoding for Qwen3.8 27B and Flash Next on one RTX PRO 6000 Blackwell: agent tasks in 36% less time and decode 2.1x on the 27B, Flash Next 1.8x on one or two GPUs, and the 374 GPU-hours it took.

Qwen3.8-27B sometimes repeated one reasoning plan for dozens of tool turns. Deleting earlier reasoning after a repeat trigger passed all 44 selected replay continuations.

Seven AI reviews of SleepFM consolidated: consensus strengths, critical weaknesses and a recommended interpretation.

Reading the SleepFM paper (Nature Medicine, 2026) with parallel reviews from several frontier models, then consolidating them.

A draft position paper: venues should engineer enforceable data boundaries and auditable AI toolchains, and enforce against harms rather than tool use.

Venue and publisher policies, empirical findings and the likely trajectory of AI in scientific peer review, as of January 2026.