NVIDIA's looped Dory, Llmspeed, MOSS in 8 steps, Drama 3 faster

A quieter day, with four small primary pages worth a look. Every number below comes from the page it links to, and I did not run any of it.

NVIDIA's Dory loops a shared block

NVIDIA released Nemotron-Labs Dory1.0-4B on 16 October. It is a 4B model with up to 256K tokens of context, under the OpenMDW 1.1 license. The design is a looped, latent-recurrent one: the layer stack has 8 input layers, 8 output layers and a 20-layer recurrent block in the middle, and that shared block can be applied several times per token. More loops means more compute per token and, on the card's table, better scores, with no retraining and no model swap.

The numbers show where it pays. On NVIDIA's own runs, SWE-bench Verified goes from 50.3 at one loop to 59.5 at two and 61.8 at four. Terminal-Bench 2.1 moves from 12.5 to 18.1 to 22.2. Most of the gain arrives at the second loop, which is why the card recommends two as the default. At four loops it edges past the Qwen3.5-9B column on SWE-bench Verified (61.8 against 61) and trails it on Terminal-Bench 2.1 (22.2 against 24), while the larger Qwen3.6-35B-A3B still leads on those two rows.

Two details matter for anyone serving it. The card says last-loop KV caching removes the extra cache memory that looping normally costs, and that looping needs fewer generated tokens per task. vLLM support is still in a pull request, so for now you install from that branch.

Llmspeed

Llmspeed lets you pick a GPU, Mac or CPU and a model, then watches the answer stream at the predicted speed, including the wait for the first token. It covers 30 devices and 15 models, has a race mode for two setups, and shows whether a model fits at each quantization. Each estimate is labeled measured, derived or projected, which is the honest part: you can see how much to trust the number before buying anything.

MOSS-Image in 8 steps

MOSS-Image was in my last issue. This is a community LoRA from lodestones that distills MOSS-Image-SFT from 50 steps at CFG 3 down to 8 network evaluations with no CFG, using a data-space DMD method. The author labels it work in progress and tells you to load the LoRA in ComfyUI at strength 1.

Drama 3 gets faster in ComfyUI

The ComfyUI-Drama3 pack says it has been about 3x faster since 17 October, using its own CUDA kernels and CUDA graphs. On an RTX 4090 with a 16 GB expert cache, a 9 second line goes from 13.6 s to 3.9 s with W4A8 weights. The pack says an 8 GB card keeps pace with real time.

What I'd watch next

Whether Dory's loop dial holds up in independent tests, and when vLLM merges support so it is easy to serve.