Community day: testing community DeepSeek V4.1 Flash recipes at home
The author reruns community-built recipes on his own hardware, including Mia's V4.1 Flash setup, a four-Spark ring, and Project Maya on a single RTX 5090, posting numbers then results.
I'm turning today into a community day 🤗😁 All day I'm posting my tests of recipes and quants that people here built and shared. Each one ran at home on my own hardware, the way its creator runs it: @0xbakeer's TandemLLM with Ornith-1.5-122B on one DGX Spark. Thanks for pointing me to TandemLLM and for answering so fast and detailed. @MiaAI_lab's Ornith-1.5-122B and DeepSeek-V4.1-Flash recipes. Mia ships bangers daily, and they run out of the box. Her recipes give me the fewest headaches when testing. @0xWhiteMage's ThinkingCap 27B quant on two engines. Thanks for asking me to test it and for the close back-and-forth. @bertholomusai's 4-Spark DeepSeek-V4.1-Flash ring and the music machine it built. Thanks for the exchange, the great recipes and the music machine idea. I'll keep building on it. Project Maya with DeepSeek-V4.1-Flash on one RTX 5090. Impressive how much it gets out of one card. But is it really that much better than Qwen3.8-27B? Flint 40B on three machines. We finally achieved AGI 🙃 Every run gets two posts: first the numbers, then the world the model built as an agent. Thanks to everyone who shares their work. 🙏