Sanasol (@rusanasol) header
Sanasol (@rusanasol) profile picture

Bearded PHP-magician

Belgrade
Joined May 2011
Photo by Sanasol (@rusanasol)Photo by Sanasol (@rusanasol)Photo by Sanasol (@rusanasol)Photo by Sanasol (@rusanasol)Photo by Sanasol (@rusanasol)Photo by Sanasol (@rusanasol)Photo by Sanasol (@rusanasol)Photo by Sanasol (@rusanasol)Photo by Sanasol (@rusanasol)Photo by Sanasol (@rusanasol)
DeepSeek V4.1 Flash Orca, now packed for Sushi by @Beamsters1 🍣 Tested on a 48GB Mac. Tiny refusal pilot: both judges marked 9/12 answers as fulfilled. Orca used 512 output tokens; comparators used 256. Full results and caveats in the model card—not a full benchmark. huggingface.co/sanasol2008/D…
37
DeepSeek V4.1 Flash Heretic, packed for Sushi by @Beamsters1 🍣 Tested on a 48GB Mac: 5–7 tok/s xD Tiny refusal pilot: judges marked 8–11/12 Heretic answers as fulfilled vs 0/12 for stock DS FLASH Maya/DS FLASH API. Caveats in the model card — not a full benchmark. huggingface.co/sanasol2008/D…
1
63
one RTX 5090, two RTX 3090s, a Ryzen 9 9950X Qwen3.8-Flash-Next IQ3_S (SC117 abliterated GGUF) on Strata DeepSeek-V4.1-Flash Maya-XS on Project Maya
162
One copy of the experts in RAM, three engines on it: RTX 5090 + two 3090s (one on OCuLink x4) all serve Qwen3.8-Flash-Next from a 50 GB shared arena. 4 streams, 371-392 t/s aggregate, 512K context on the 5090. Gotcha: /dev/shm quota kills it silently at 36 GiB.
1
33
Our two used Chinese RTX 3090s screamed 24/7 — 3300 RPM even at idle, even at 16 W. We couldn't flash a fix: Ampere rejects edited VBIOS. Now they idle at 0 RPM, and the flash chip was never touched. Thread on the trick nobody tells you about 🧵
1
97
For the brave: Ampere fan tables live beyond the 512K ROM window (full PROM read via BAR0+0x300000, driver unbound). The min_* fields are read-only in editors → hand-patch u16s, fix the additive checksum, normalize to ampere-relocated-v1 (512K GSP cap).
1
79
My first local 1B tokens, took 1 week of relaxed work Qwen3.8 Flash Next and Qwen3.8 27b (retired now)
22
A Steam Deck OLED running a 27B LLM server on an RTX 5090 — 318 t/s decode on its own screen. M.2 → OCuLink → eGPU dock, PCIe 3.0 x4, Ubuntu on microSD — and decode matches full-desktop benchmarks. ~€80 of adapters. The whole stopgap build, pitfalls included 🧵
3
1
4
532
@OnDeck @Hi_MINISFORUM — your dock + your handheld served a 27B LLM for days 🙂
11