TensorFold Beats vLLM on Two DGX Sparks by 2.26x on Chat

@johnennis shares the agent trick he calls phenomenally powerful. Ask the model for an HTML page of its work, click to comment on the areas, then copy the comments back into the session. He storyboards videos this way, with each comment stamped to the area it describes. @TheBalkanHacker drops such a page into every UI project for direct feedback, and @jetpen had been building the same thing as a skill before concluding he does not need to.

@sudoingX loaded NVIDIA’s NVFP4 GLM-5.3 Flash on two DGX Sparks with stock vLLM, 256GB of unified memory across the two boxes and weights split over the ConnectX cable. His numbers: 320B total parameters, 18B active, 204GB of weights; decode of 15 tok/s with MTP off, flat to 128K, and 19.5 to 31.7 tok/s with MTP on at k=3. A 128K prompt returned its first token in 84 seconds. He says quality testing and tuning come next.

@sudoingX describes the setup he calls the most useful thing in his local AI stack. Two DGX Sparks run one vLLM endpoint behind Tailscale, so every device he owns reaches the same model with no port open to the internet. The server speaks OpenAI chat, OpenAI Responses and Anthropic Messages, and answers to the name “local”. Swapping GLM 5.3-Flash for another model then requires no client config change, and requests queue when several devices fire at once.

@0xBakeer is running dense Qwen3.8-27B, not Qwen Flash, on a single DGX Spark with his TandemLLM engine. Optimized caching and the StairCut method bring a roughly 150k-token prefill down to milliseconds, and decode hits around 50 to 90 tok/s depending on the task. In a clip recorded after five hours of continuous running, he had Qwen design a pagoda garden and keep iterating on new ideas without stopping.

@techNmak calls “the model fits in VRAM” one of the most misleading sentences in LLM engineering. Weights are the predictable part of the footprint; each active sequence adds KV state that grows with retained tokens. In his worked example, eight concurrent requests retaining 16,384 tokens each need 16 GiB of KV cache against 14.90 GiB of weights, 30.90 GiB together, beyond a 24 GiB GPU before activations. His remedies: grouped-query attention, PagedAttention, prefix caching and KV-cache quantization.

Clinton Donnelly, @CryptoTaxFixer, believes the IRS has sharply increased Letter 6174 mailings to crypto traders and estimates about 10,000 a month, after five clients brought six letters in one week. He describes 6174 as a voluntary compliance letter, not an audit, and notes some versions state no response is required. His advice: get professional advice before signing or sending anything. @InsomniaLK calls the volume brutal; Donnelly answers that a 6174 does not automatically mean an amended return.

Pirate Face, a new site for what it calls sovereign AI, mirrors eligible Hugging Face models as torrents and records each weight file’s SHA-256 for comparison against the source. If a model disappears from Hugging Face, the magnet keeps working while peers seed, and recovered models are marked Rescued. A drop-in API is planned that swaps HF_ENDPOINT to pirateface.co. Eligible MIT and Apache-2.0 submissions list immediately; Kimi-K3 and Audn’s Penclaw GLM-5.3 carry model-specific exceptions.