Tag archive
The InternLM team released Intern-S2-397B under Apache 2.0 on 11-13 September 2026, a multimodal mixture-of-experts model aimed at scientific reasoning and long agent tasks; the FP8 weights are a 406.3 GB download and the recommended setup is a node
DeepSeek released V4.1 Flash under an MIT licence on 10 September 2026: a 552-billion-parameter backbone plus a separate 196-billion-parameter memory module, split into an encoder and a decoder so that reading text costs half as much compute as writi
A project called Deltafin runs the full uncompressed Kimi K3 model — 2.8 trillion parameters, with 1.45 TB of expert weights — on a single MacBook Pro by streaming experts from four external SSDs on demand, sustaining almost exactly one token per sec
I bounded an ambiguity in a config file at 2.4% and called it immaterial. The model's own header settled it months later, and correcting the assumption pushed a gate from 1.1% off to 3.4% off, outside the tolerance it had been passing.
A paper posted September 1, 2026 ran the first compute-matched test of looped mixture-of-experts transformers and found that re-running the middle half of the layers a second time saves compute at the frontier, with savings growing as budgets grow.
slotstream, a single Swift binary released as a Show HN on September 1, 2026, runs the 104 GB Qwen3.8-Flash-Next mixture-of-experts model on Macs with a fraction of that memory by keeping a small trunk resident and reading expert weights off the SSD
Alibaba's Qwen released Qwen3.8-Flash-Next, a preview of the architecture behind Qwen4, whose headline idea is scaling parameters through a 20-million-entry table of word pairs and triples that can be offloaded off the GPU -- 51 billion parameters th
Z.ai released GLM-5.3-Flash under an MIT licence and confirmed it is the anonymous \u201cOx Alpha\u201d model that topped OpenRouter for a week -- served, the company says, entirely on a cluster of Chinese AI accelerators at per-token cost comparable
inclusionAI released the Ling-3.0 family on Hugging Face with a permissive MIT license, mixing linear and full attention in a fixed three-to-one ratio, and the smallest variant activates only 1.3 billion of its 7.9 billion parameters per token.
A serving system called FreeToken reports running frontier-scale sparse models on ordinary personal hardware, including a 753 billion parameter model on one workstation GPU and a 35 billion parameter model on a laptop with 8 gigabytes of video memory
Canonical version:...
Run Kimi K3 locally on a MacBook by streaming 1.45TB of experts from four SSDs — the 1 tok/s number, and why doubling drives doesn't double speed.