TensorFold 1.0.7 has landed with a major overhaul to Living Weights. • Each fact you teach through /learn becomes about ten new neurons inside the model's own weights, learned about 4x faster, with a live 3D view of what it has learned. Macs and DGX Spark. • Structured output and required tool calls on the native server, with drafting still on. • Qwen3.8-27B serves concurrent requests, and sampled replies draft too: 125 tok/s vs 20 on code on an M3 Ultra. • RTX 50-series and RTX PRO 6000 Blackwell support, NVFP4 Nemotron on DGX Spark, and faster GLM-5.3-Flash on two Macs. • Various other bugs and issues squashed I would like to thank the people who reached out to me and gave me access to their hardware to speed up testing. 🫶🏼 Thank you to everyone who sent PRs, reports and receipts, and to everyone using it: BobClawblaw BlivionIaG jschmied ZackSample ThinkOffApp jayleaton chaog992 Nipale-ai @volatilemarkts @MiaAI_lab @petruspennanen @plotarmordev

TensorFold 1.0.3 lands with a few cool things. - Living Weights (beta with Nemotron on Metal) and coming to CUDA in 1.0.4 - Nvidia Ampere (beta) for Nemotron, and native Qwen3.8-27B with DFlash2 - Conversation prefix caching - Various bugs fixed and improvements Twelve contributor PRs helped make this happen, so a massive thank you to @niklaslenz_ai and GitHub contributors akol1, jschmied, chaog992, AdrianBinDC and BobClawblaw, and everyone testing, reporting bugs and helping us improve TensorFold. FYI, Living Weights is in very early beta and has sharp edges, but I think I’m on to something. I’d love community feedback to make it better: @plotarmordev @petruspennanen @MiaAI_lab @volatilemarkts

Ran TensorFold (EXL3) against my own vLLM TP2 recipe, same prompts: ⚡ Code 82 vs 69 tok/s ⚡ JSON 84 vs 71 ⚡ Prose 44 vs 35 📈 Mixed load at C4: 108 vs 72 tok/s (+51%) 🚀 Prefill 1,781 vs 1,225 tok/s 🧠 500K context vs 262K Two boxes on a desk 🤯

New Tensorfold Recipe I am working on for GLM 5.3 Flash. NVFP4 and EXL3 Lanes, going to be working and tuning this though more. I still seem to not be getting better numbers than vLLM.

Part 4 of my TensorFold 1.0 test: long prompts. Does @ashxhart's new native engine hold up at 128k tokens? Same GB10 box (Lenovo PGX), Nemotron 3.5 Lightning, real book text from Project Gutenberg, old Python engine 0.6.6 vs 1.0.1, median of 3 different book passages per length: decode at 8k: 118.7 → 138.0 tok/s decode at 32k: 110.5 → 124.0 tok/s decode at 64k: 97.5 → 111.8 tok/s decode at 128k: 85.9 → 96.2 tok/s So 1.0.1 decodes 12–16% faster at every length (short prompts in part 1: 1–9%). Prefill is a tie: 0.6.6 is up to 2% faster (a 128k prompt takes 31.8 s vs 32.3 s to the first token). And all 12 answers to these long prompts are bit-identical between the two engines, from 8k to 128k. One limit: with 4 long prompts (32k each) at once, both engines read them one after another. The 4th request waits about 22–24 s for its first token on either engine, and all 4 finish in about 24 s.

Lots of you asked for a TensorFold Discord, so it's finally here. Come share model recipes and benchmarks, report bugs, get setup help or join the contributor discussions. Come say hello: @MiaAI_lab @volatilemarkts @petruspennanen @plotarmordev

🚀 DeepSeek V4.1 Flash 1017 now runs on TensorFold across 4 DGX Sparks ⚡️ 89.8 tok/s for 1 user (vLLM: 48.5) 🔥 202.9 tok/s with 6 users 🧠 LibertAI's 3.5bpw EXL3, every matrix 🛠️ No new kernels needed New default lane in the repo 👇