DWS Research

DWS Research is where DWS works on problems it doesn't know how to solve yet.

The work is open-ended on purpose, so projects aren't tied to a product or a ship date. A project runs until its question is answered or clearly won't be, and the numbers get written down either way.

Research is what I'm doing when I don't know what I'm doing.

Fig. 1 "Research is what I'm doing when I don't know what I'm doing," attributed to Wernher von Braun, split by DWSLM's tokenizer into 16 tokens. "Research" alone takes two.

01Focus areas.

machine learning
DWSLM, a 1B-parameter language model built from scratch
computational photography
Amidol, film simulation developed on the phone from one RAW frame
computational audio
Skiff, Audio Unit chains on a Mac's output and microphone
distributed systems
the DWS network, BGP anycast across five countries
high-performance computing
nothing public yet
site reliability
public DNS and NTP on that network

Most projects stay private until they're ready to be read. The ones with public pages are linked above.

02DWSLM, in progress.

dwslma 1B language model, from scratchmodel dwslm/data15.7M documents, 23.2B tokensdone dwslm/tokenizerDWSLMTok32.3, 32,768 tokensdone dwslm/model256M first: proxy runs and main run donedone
dwslm/256mDWSLM-256M released: base weights on Hugging Face, chat demo, model cardreleased
dwslm/kernelsday 0: every kernel checked and timeddone dwslm/dashboardevery training run, livelive

DWSLM is a 1B-parameter language model with nothing borrowed. The corpus, tokenizer and training code are ours, and training will run on a single GPU. Its pages follow each phase as it finishes, including the parts that didn't work.

DWSLM →

parameters, target
1B
training tokens
23.2B
open sources
18
vocabulary
32,768

data, tokenizer, model: donepretrain: now

Fig. 2 DWSLM as of 2026-11-11.

03How we work.

  • We write the parts that decide a result instead of importing them.
  • Every number on these pages is measured, or labeled as an estimate.
  • The hardware stays small on purpose. A result that needs a cluster is someone else's to find.
  • Retired versions stay on the page, with the reason they were retired.

DWSLMTok32.1, plain BPE · 30 tokens

研究就是在不知道自己在做什么的时候所做的事。

DWSLMTok32.3 · 22 tokens

研究就是在不知道自己在做什么的时候所做的事。

Fig. 3 A retired version. The first tokenizer split common Chinese characters into byte fragments, marked here with small numbers. The fix is on the tokenizer page.

04Same stack as the rest of DWS.

DWS runs its own network, with its own ASN and IP space and BGP anycast across five countries. Public DNS and NTP run on top of it, and none of it sits in a public cloud.

Research runs on DWS hardware too, with our own code and openly licensed data. This site is served from the same anycast edges as everything else at dws.rip.

asn
AS30179
ipv6
2602:f4e8::/40
ipv4
157.254.184.0/24

Questions, corrections and collaborations go to admin@dws.rip.