@gptproto-jacob.bsky.socialMicrosoft demoed DeepSeek V4 Flash running on a laptop:
284B open-weight, quantized to 1.6 bits → ~60GB.
At 8 bits the same weights need ~284GB.
Memory, not compute, decides if it runs.
On-device inference doesn't kill the bill —
it changes its unit. #LocalInference
@nile11.bsky.socialMicrosoft Surface Laptop Ultra: 128GB RAM for AI Devs
..................
https://nile1.com/microsoft-surface-laptop-ultra-128gb-ram-for-ai-devs/
..................
#128gbRam #aiDevelopers #largeLanguageModels #localInference #macbookPro #microsoft #surfaceLaptopStudio #surfaceLaptopUltra #unifiedMem
@futuregearai.bsky.socialA new arXiv paper, HARISSA, uses a language model's own hidden states to decide which queries deserve extra compute and which are too risky to answer, helping small local models balance efficiency and safety without sending…
#EdgeAI #OnDeviceAI #LLM #LocalInference
https://arxiv.org/abs/2609.38006
@ostechnix.bsky.socialLearn how to run a private, offline AI coding assistant with Atomic Chat on Linux. Install it, download a coding model, and code locally without cloud API keys.
Full guide here: ostechnix.com/atomic-chat-...
#AtomicChat #AI #CodingAgent #AIAgent #Opensource #Linux #Windows #Macos #LocalInference