TensorFold 1.0.7 writes each learned fact into ~10 new neurons, 4x faster with 3D view
HankYeomans · x · 2026-10-11
TensorFold 1.0.7 overhauls its Living Weights system: each fact taught via /learn becomes about ten new neurons inside the model's own shards (a 64-neuron pocket per MoE shared expert), learned 4x faster with a live 3D view, on Macs and DGX Spark. Facts are only kept when held-out wordings recall them without side effects; no adapters, readable by any MLX loader. Also adds structured output and required tool calls, concurrent Qwen3.8-27B serving at 125 tok/s on M3 Ultra (vs 20), RTX 50-series and RTX PRO 6000 Blackwell support, NVFP4 Nemotron on DGX Spark, and faster GLM-5.3-Flash across two Macs.
Related event: TensorFold Ships Living Weights for Continually Learning Local Models(2 posts)→
More from Infra
- Report: Anthropic signs $11.6B, 7-year compute deal with Akamai — nikola_mr64990 · 2026-10-11
- Musk confirms Terafab will modify existing tools and build custom equipment to lift wafer output — elonmusk · 2026-10-11
- Musk says RAM supply is the problem as quantized 25GB MoE local run speculated for 2028-29 — elonmusk · 2026-10-11
- Quartermaster: open-source local server runs Qwen-Image 2.1 Turbo without ComfyUI — OneMoreName1 · 2026-10-11
- Nvidia backs $105B of OpenAI data center rent; Google guarantees $50B+, mostly for Anthropic — AccBalanced · 2026-10-11
- IEEE sees developed-country energy storage doubling in 2-3 years as tech infrastructure hits ~10% of global electricity — ingliguori · 2026-10-11