X / @superalesha · Hugging Face · alesha.pro · t.me/fuckup_files

By day I run infrastructure. The rest of the time I bench LLMs on a hand-built 4x RTX 3090 rig with 96 GB of VRAM, and I publish every measurement with raw traces and per-run provenance attached. People argue about quants and reasoning effort from vibes. I run the thing and count.

Project	Stars	What it is
tools		Things I build while running models on my own hardware: ComfyUI workflows, model setups, agent skills and benchmark configs from the rig
atlas		Interactive canvas that takes an LLM apart tensor by tensor. It measures INT8/INT4/FP8 error, distributions, spectra and outlier channels on every weight tensor
bench-portal		A catalog of games, 3D showcases and shader experiments built by AI models, filterable by the model that built them. Live
llama.cpp-mirai-s		llama.cpp with Mirai S support: a 2.4-bit Qwen3.8-27B in 12 GB of VRAM with 128K context, plus refusal removal at run time
qwen38-flash-next-4x3090		Qwen3.8 Flash Next AWQ INT4 on 4x RTX 3090: 262K context, vision, FP8 KV cache and CUDA Graphs
redcast		An agent-native, Raycast-style launcher for macOS. My fork of Tinycast with native and ACP agents, voice and screen awareness
4x3090-llm-benchmarks		Public archive of local LLM speed measurements from the rig, with exact launch commands and per-run provenance

My quants, abliterations and engine builds live at huggingface.co/alesha-pro.

• Qwen3.8 Flash Next Abliterated: EXL3 and Strata GGUF: the uncensored Qwen3.8 Flash-Next for Strata and exllamav3, over 220,000 downloads.
• GLM-5.3-Flash on 4x RTX 3090: GLM-5.3-Flash quants and the abliterated Maya build that fit four 24 GB cards.
• Qwen3.8-27B-S-mirai-GGUF: Mirai S in 12 GB of VRAM, with the vision projector and the refusal vector.

The rig is 4x RTX 3090 with 96 GB of VRAM in total. P2P is verified on all 12 pairs, the cards run at 220 W each on PCIe 3.0 x16. Engines on the box: vLLM, llama.cpp, exllamav3 + tabbyAPI, plus SGLang when it behaves. Ampere sm_86 has no native FP8, and most quant advice out there assumes everyone owns a Hopper. I enjoy measuring what actually wins on these cards.

• NVIDIA sold millions of 3090s with half the tensor speed switched off. One kernel module took my BF16 tensor pipe from 82 to 155 TFLOPS, with outputs matching the stock run byte for byte.
• DeepSeek just published the price of a token. 155 trillion tokens a day at $38 of revenue per GPU-hour: at these prices, serving a frontier-class model is a very profitable business.
• 110,409 downloads in 9 days for the uncensored Qwen3.8 Flash-Next I put up for Strata. My part was a 480 KB file and a weekend of measuring.
• A 27B model in 12 GB of VRAM. Mirai S Qwen3.8-27B at 2.4 bits per weight: full 262K context, 70 tok/s, and now uncensored.
• Local AI performance is a software problem. A 2020 RTX 3090 runs Qwen3.8 27B up to 4.6x faster, and nobody bought a thing.
• My RTX 3090 made these while I was asleep. 103 images and 63 clips overnight with Qwen-Image-2.1 and MiniMax H3. No cloud, no credits, no queue.

Every number on this profile comes from a real run. If it is not in a repo with raw data, I did not say it.