Real-world
Research & analysis
Concrete workflows, including the tradeoffs and limitations reported by their authors.
this collection
Pre-digesting 1,500 companies' reports before a hosted model sees them
A local model reads financial reports across 1,500 companies first, so the more expensive hosted model only has to work with the distilled output.
Academic and policy research on an M5 and a DGX Spark
Purpose-built tools around a personal research corpus, plus Hermes as an interactive research assistant, get this researcher roughly 80% of the way to frontier quality — locally.
Local only for what is deterministic: embeddings, speech and TTS
After testing every model that fits a 128GB Strix Halo, this user keeps local strictly to embeddings, text-to-speech and speech-to-text, and relies on cloud for reasoning.
A 128GB Strix Halo research stack, with a warning against smaller machines
A researcher runs Qwen 3.8 Flash Next and Gemma 4 31B on a 128GB Strix Halo, splitting strategy and summarisation across models and advising against 64GB machines.
Qwen 3.8 Flash Next on 12GB of VRAM via RAM+VRAM offload (Strata engine)
Four quants of a vision model measured on a 12GB RTX 5070 SFF with 64GB DDR5, through a purpose-built inference engine, at 37 to 55 GB combined RAM+VRAM.
Running Qwen3.8 Flash Next 176B on a 16GB RTX 3080 Laptop with 32GB RAM and SSD
TensorSharp runs the 176B Qwen 3.8 Flash Next MoE on a 16GB RTX 3080 Laptop with 32GB RAM, matching Strata on decode while finishing the benchmark process far faster.
Benchmarking Strata n-gram table placement on a modded RTX 4090 48GB with 128GB DDR5
A measured benchmark on a modded 48GB RTX 4090 shows Strata's n-gram table belongs on SSD while conversation parking delivers the large speedup.
Turning an hour of stock research per ticker into minutes
Local models take over the reading and summarizing side of manual stock analysis, cutting a research routine from hours a day to a few hours a week.
A local Deep Research replacement that runs on a few dollars a month
A local deep-research pipeline using GPT-OSS 120B and Gemma 12B produces results on par with cloud deep research at a fraction of the token cost.
GLM 5.3 Flash and MiMo-V2.6-Flash on one 128 GB Strix Halo mini PC with the Kyojin engine
Release benchmark of an ExLlamaV3-based ROCm engine running two 300B-class MoE models, each on a single 128 GB Ryzen AI Max+ 395 mini PC.
Dual Radeon Instinct MI50 benchmarks: dense models at 12 to 18 tok/s, MoE models at 47 to 61 tok/s
Two Radeon Instinct MI50 cards limited to 145W each are benchmarked with llama.cpp Vulkan across dense and MoE models, with MoE models far ahead of dense ones.