Ali Asif

The AI log, agents and coding · science

ComfyUI doubles local LLM speed, Google's CKM health age, Tabletop

A 2x local token-rate gain in ComfyUI, Google Research on wearable health age and coaching, and the Tabletop group-planning experiment.

By , published

What shipped

  1. ComfyUI

    Doubling LLM token rates in ComfyUI

    A rewrite of the text generation hot path for Qwen3.x checkpoints lifts Qwen3.5 9B on an RTX 5090 from 106 to 219 tokens per second, shipped in ComfyUI v0.45.0.

  2. Google Research

    Using wearables to enable cardiovascular-kidney-metabolic (CKM) risk profiling for personalized coaching

    An investigational framework that turns wearable data into a coachable health age and simulates 12-week habit changes with a world model; coaches preferred its plans in 71.4% of evaluations.

  3. Google Labs

    Tabletop: Collaborative group planning with Google Labs

    A new Labs experiment that generates shared interactive itineraries, polls, expense splitters and brackets from links and prompts, rolling out by waitlist in the U.S.

A quieter stretch today, so this is three items rather than a long list. One is a plain speed gain for people running models locally, one is a health research paper with real numbers attached, and one is a small product experiment from Google Labs.

ComfyUI doubles local text generation speed

ComfyUI runs language models locally for prompt enhancement, image description and story writing. The team rewrote the hot path of that text generation, focused on the Qwen3.x family, and says the same checkpoints on the same GPU now generate tokens up to twice as fast, with no change to the model or its output quality.

The figures in the post: Qwen3.5 9B went from 106 to 219 tokens per second on an RTX 5090, and from 91 to 129 on an RTX 5080. Qwen3.8 27B describing a 1536 by 1536 image went from 55 to 122 tokens per second on the 5090. A W4A8 build of the 27B model reaches 137 tokens per second with the mtp option left on auto.

The explanation is useful on its own. A single workflow generates one sequence, so every step is a batch of one and the limit is how fast the GPU can read the weights from VRAM. For a 27B model at four bits that is roughly 14 GB read per token. Diffusion models are bound by compute instead, which is why the two workloads need different tuning. It is all in ComfyUI v0.45.0: load a Qwen3.x checkpoint and use the Text Generate node.

Google Research on a coachable health age

Google Research describes an investigational framework for cardiovascular-kidney-metabolic risk. It builds a health age from modifiable daily behaviour and physiology captured by consumer wearables, then uses a world model called SensorWM, trained on over 2.2 million person-weeks of data, to simulate 12-week what-if habit changes. A language model turns the top-ranked paths into a coaching plan.

In a blinded comparison with five certified lifestyle coaches and five physicians across 49 participant profiles, coaches picked the counterfactual-grounded plan as best overall in 71.4% of evaluations. Physicians rated 99.2% of plans within safe exercise limits. A companion study adds three routine blood biomarkers for 819 participants and finds that early kidney decline often shows up independently of metabolic and cardiovascular strain.

Tabletop from Google Labs

Tabletop is a Labs experiment for groups. You add links, files and ideas, then prompt it or pick a template to generate a trip itinerary, dinner poll, expense splitter or tournament bracket that everyone can edit at once through an invite link. It rolls out in the U.S. for ages 18 and up via waitlist.

What I'd watch next

Whether the ComfyUI gains hold on other model families, and whether the SensorWM approach gets evaluated outside the group that built it.

Sources

Topics: NVIDIA, Qwen, Google Research, World Models, ComfyUI