DeepSeek-V4-Flash-0731 (all 284B parameters) running on a single DGX Spark. ⚡ ~17 tok/s generation, 40 tok/s prefill 📦 One 80GB GGUF file, mixed IQ2_XXS/Q8 imatrix quant 🧠 256-expert MoE, 6 routed per token 🆓 MIT licensed frontier model (kinda) on your desk 😎 🤗

DeepSeek v4 Flash is actually pretty good. Is it as good as the benchmarks claim? No. It is benchmaxxed like everything else shipping right now. But it is a 284B parameter model. At that size it has no business performing this well. It beats GPT 5.6 Luna in real use. A model a fraction of the size, from a lab a fraction of the budget. DeepSeek keeps doing more with less than anyone in the world. Very impressed.

DeepSeek switched off 2 of its model names. If your app still calls them, it's broken right now. Renaming takes ten seconds. The replacement has "thinking" on by default, so you quietly pay more per call. Nobody notices until the invoice.

DEEPSEEK PAUSES FUNDRAISING Chinese AI startup DeepSeek has reportedly paused its second fundraising round, informing prospective investors that no investment agreements will be signed for now, according to Bloomberg. The company may restart the process later. Reuters previously reported DeepSeek was also exploring a potential IPO on Shanghai’s STAR Market.

DeepSeek V4 Pro costs 69× less than Fable 5 per task according to this chart. That gap may be one of the strongest arguments for open-weight AI. So how does DeepSeek make it so cheap? ↓ V4 Pro has 1.6 trillion parameters, but its MoE architecture activates only 49 billion per token. In other words: it gets the capacity of a gigantic model without running the entire model for every word. Then it cuts the inference bill further: • Sparse, compressed attention reduces long-context compute • FP4/FP8 precision lowers memory and bandwidth requirements • Prefix caching makes repeated inputs dramatically cheaper • Open weights let the entire ecosystem optimize deployment Cached API input is also priced roughly 120× lower than uncached input.