tag: #self-hosted (25 posts)

• 2026-09-25 :: A year of Fiona: my personal assistant, the Hermes swap, and the agent that rewrote its own playbook :: #agentic-ai, #ai, #llm, #self-hosted
• 2026-09-21 :: MTP on Strix Halo: +79% at one stream, zero at four, and the ceiling isn't bandwidth after all :: #ai, #hardware, #llm, #self-hosted
• 2026-09-21 :: A Strix Halo engine audit: a scary claim that didn't reproduce, a fork worth +35-92% prefill, and seven broken routes :: #ai, #hardware, #llm, #self-hosted
• 2026-09-17 :: DwarfStar, DiffusionGemma, and the ROCm 10 unfreeze: an ecosystem status report from halo :: #ai, #llm, #self-hosted
• 2026-09-17 :: Four coding models on one Strix Halo: pass@1 hides two very different failure modes :: #agentic-ai, #ai, #hardware, #llm, #self-hosted
• 2026-09-17 :: MTP on halo: two production adoptions, one banned backend, one corrected claim :: #ai, #hardware, #llm, #self-hosted
• 2026-09-17 :: Do synthetic LLM benchmarks lie? On halo, real code matched llama-bench to 0.6% :: #ai, #hardware, #llm, #self-hosted
• 2026-09-16 :: Cerebras vs our H100: 11x the decode speed, half the agent turns :: #agentic-ai, #ai, #llm, #self-hosted
• 2026-09-15 :: An afternoon of agentic coding: 12,53 € on our H100, 47 € to 726 € through an API :: #agentic-ai, #ai, #llm, #self-hosted
• 2026-09-15 :: LFM2.5-2.6B, the silent submarine, and why retrieval doesn't judge what it retrieves :: #agentic-ai, #ai, #llm, #self-hosted
• 2026-08-31 :: Qwen3.8-27B-INT4 with MTP: 30-40% faster round trips, and two rival explanations for the divergence :: #agentic-ai, #ai, #hardware, #llm, #self-hosted
• 2026-08-30 :: Qwen3.8-Flash-Next on Strix Halo, part 2: the missing speed was already merged upstream :: #ai, #hardware, #llm, #self-hosted
• 2026-08-28 :: Qwen3.6-35B-A3B on Strix Halo: one flag buys 22% prefill, and my own GPU power advice was wrong :: #ai, #hardware, #llm, #self-hosted
• 2026-08-28 :: strix-halo-optimize: a tuning skill whose rules had to become code :: #agentic-ai, #ai, #hardware, #llm, #self-hosted
• 2026-08-27 :: Qwen3.8-Flash-Next on Strix Halo: a 177B Qwen4 preview at 22 tok/s, one day after release :: #ai, #hardware, #llm, #self-hosted
• 2026-08-21 :: MTP speculative decoding costs code correctness on two Qwen-lineage models :: #agentic-ai, #ai, #llm, #self-hosted
• 2026-08-21 :: Ornith 1.5 35B A3B on Strix Halo: MTP pays off at n=1, not at the default n=3 :: #agentic-ai, #ai, #hardware, #llm, #self-hosted
• 2026-08-12 :: The small model didn't fabricate. It stopped citing. :: #agentic-ai, #ai, #llm, #self-hosted
• 2026-08-07 :: AMD buys Taalas: etched models, and what it means for local inference :: #agentic-ai, #ai, #hardware, #llm, #self-hosted
• 2026-07-27 :: Poolside says Laguna needs an H200. We ran it on a single H100. :: #agentic-ai, #ai, #hardware, #llm, #self-hosted
• 2026-07-26 :: DeepSeek-V4-Flash on Strix Halo: it runs, and now we know how fast :: #ai, #hardware, #llm, #self-hosted
• 2026-07-26 :: Laguna-S-2.1 on a mini-PC: the honest numbers :: #agentic-ai, #ai, #hardware, #llm, #self-hosted
• 2026-06-15 :: Running an H100 at Trail Openers: what it actually costs in money, energy, and CO₂ :: #agentic-ai, #ai, #hardware, #llm, #self-hosted, #sustainability
• 2026-06-05 :: H100 vs Strix Halo: the gap is bigger than the first benchmark suggested :: #agentic-ai, #ai, #hardware, #llm, #self-hosted, #sustainability
• 2026-04-30 :: New Strix Halo? Five things that will cost you hours. :: #ai, #hardware, #llm, #self-hosted