Compiled for History's Future: The Singularity Is Here — curated from Hugging Face trending models, datasets, and research papers.
Field Pulse
DeepSeek-V4.1-Flash dominates this week by a margin rarely seen on the Hub: a trending score of 1,942 — more than double the runner-up — on 244,500 downloads and 2,147 likes within days of its September 10 release. It extends the "Flash" efficiency race identified two weeks ago (three major labs shipping frontier-class-at-lower-cost models in the same window), and this week DeepSeek's entry is decisively the loudest signal in the field. Efficiency is no longer a secondary concern behind raw capability — it is now the primary axis labs compete on.
The more structurally interesting story is at the edges. Edge0's Edge0-35B-A3B-preview, built natively for Apple silicon via the mlx library, combines a Mixture-of-Experts "prerouter" with SSD-offload — paging inactive experts from disk rather than holding a full 35B model in memory. It has drawn 826 likes on just 3,600 downloads, an unusually high like-to-download ratio that signals technical excitement from people evaluating the technique itself, not casual users. A second new lab, nex-agi, debuted two multimodal MoE models (Nex-N2.5-mini and Nex-N2.5-Pro) the same week — evidence that the barrier to entry for shipping a credible frontier-adjacent lab keeps falling.
Underneath both stories sits AuK, an open-source unified speech generation and editing model that collected 217 upvotes — the highest of any paper this week by a wide margin — for unifying speech synthesis and editing under natural-language instruction. Pair that with Show-Harness, which links vision-language models to robot control through discrete semantic actions, and the pattern completes itself: intelligence is compressing (Flash models), spreading to new hardware tiers (on-device MoE via SSD offload), and reaching into physical embodiment (VLM agents operating robots). None of these are separate trends. They are the same trend, viewed from three different angles.
This week's runaway leader with a trending score of 1,942 — more than double the nearest competitor. Released September 10, it extends DeepSeek's V4 line into the Flash-efficiency race that three major labs entered simultaneously two weeks ago, and does so more decisively than any single release since.
A new lab's debut model, built natively for Apple silicon via mlx, combining an MoE "prerouter" with SSD-offload to run large models by paging inactive experts from disk. 826 likes on only 3,600 downloads is an unusually high engagement ratio — a signal of technical excitement, not casual adoption.
A previously unseen lab, nex-agi, debuted two multimodal MoE models in the same week. The Pro variant targets maximum capability; its companion Mini targets efficiency — the same two-tier release playbook now standard across the field, this time from a brand-new entrant.
The efficiency-tier companion to Nex-N2.5-Pro, drawing more likes than its larger sibling despite fewer downloads — early evidence that developers are specifically seeking out the smaller, cheaper variant to evaluate first.
A compact music-generation model tagged for both symbolic planning and agentic editing — framing music composition as a multi-step planning task an agent can iterate on, rather than a single generative pass.
Still climbing three weeks after release: downloads have risen to 135,700 from 82,300 two weeks ago, alongside continued modification activity this week. The largest published corpus of recorded professional CAD, BIM, and structural-analysis workflows remains one of the field's clearest bets on AI operating inside expert software.
AuK Technical Report: An Open-Source Foundational Model for Speech Generation and Editing
This week's highest-upvoted paper by a wide margin (217), AuK unifies speech generation and editing under natural-language instruction and audio context via a multimodal language model, joint VAE, and hybrid rectified-flow transformer. Collapsing generation and editing into one instructable model — rather than separate specialist tools — mirrors the same unification pattern seen in video and world-model generation.
Show-Harness links vision-language models to robot control through discrete semantic actions interpreted by embodiment-specific modules, enabling zero-shot and efficient fine-tuned deployment across different robots and GUIs alike. The paper's real claim is that a single VLM agent, not a robot-specific control stack, can generalize across physical embodiments — the same agent architecture doing digital and physical work.
Marigold V2: Revisiting Diffusion Transformers for Monocular Depth Estimation
Marigold V2 repurposes diffusion transformers for single-image depth estimation via single-step flow-matching inference and a two-stage fine-tuning protocol, producing sharper out-of-distribution depth maps than prior approaches. Accurate depth from a single camera is a quiet but necessary building block for exactly the kind of embodied, physical-world agents this week's other releases are reaching toward.