Mitesh·5d agoLLM Decoded: Rolling the Dice on 151,936 WordsThe LM head and the sampler: how 2,560 numbers become one word, and what temperature, top-k, top-p and seed actually control.
InSparktasticbyAmegilla·6d agoBuilding a voice-based Improv Coach on DGX SparkA local AI improv coach to help me practice between classes
Mirek Fokt·6d agoSelecting a model for local agentic AI inference in SecOps on NVIDIA DGX SparkThis is my next round of NVIDIA DGX Spark testing, following NVIDIA DGX Spark on My Desk: A Builder’s Review of the Local Inference Machine…
InGenerative AIbyJim Clyde Monge·Nov 5NVIDIA Engineers Boost MoE Inference on NVIDIA DGX Spark by 42.5%How Ling-3.0-flash’s sparse MoE design helps NVIDIA deliver more useful work per token on a single DGX Spark.
Mitesh·Nov 4LLM Decoded: 36 Layers, 144 KB per Token, 30 GB of MemoryWhy Qwen is deep, why every layer keeps its own keys and values, and where your chat history actually lives while you talk: the KV cache.
Mitesh·Nov 4LLM Decoded: Where the Model Actually Thinks: the MLPThe second half of every layer, where most of the model’s knowledge lives, and the one rule for how anything gets written onto an…
YSY·Nov 3A Voice Is Not Yet a TeamEpisode 07 of Building Kigenka’s DGX AI Media Studio & Lab — Chronicles from a Human–AI Production Studio Built on One DGX SparkA response icon2A response icon2
InProfit4Cloud — AIbyIddo Kerkhof·Nov 3Pushing the DGX Spark to the Max with Qwen 3.6 and 3.8 Side by SideGitHub: https://github.com/Profit4Cloud/dgx-spark-vllm
Mitesh·Nov 3LLM Decoded: The 32 Questions Every Word AsksAttention in Qwen3: what a layer is, what a head is, what queries, keys and values do, and why 32 heads share only 8 sets of keys and…
Mitesh·Nov 1LLM Decoded: 2,560 Numbers Walk Into a ModelEmbeddings: how Qwen trades a token id for a list of 2,560 numbers, and why those numbers mean nothing one at a time and everything…
