Mobius Labs (@Mobius_Labs) header
Mobius Labs (@Mobius_Labs) profile picture

Mobius Labs

@Mobius_Labs

Multimodal AI for the world's scale. Proponents of Open Source and Open Intelligence. https://xitter.cf/t.co/1nC6r8hOrE for some of our recent work.

 Berlin, Germany
mobiuslabs.com
Joined April 2018
  • Tweets420
  • Following214
  • Followers3.3K
144Media
Photo by Mobius Labs (@Mobius_Labs)Photo by Mobius Labs (@Mobius_Labs)Photo by Mobius Labs (@Mobius_Labs)Photo by Mobius Labs (@Mobius_Labs)Photo by Mobius Labs (@Mobius_Labs)Photo by Mobius Labs (@Mobius_Labs)Photo by Mobius Labs (@Mobius_Labs)Photo by Mobius Labs (@Mobius_Labs)Photo by Mobius Labs (@Mobius_Labs)
Replying to @Dropbox
@Dropbox is acquiring our IP, and the Mobius core tech team is joining them. A huge milestone for us! dropbox.tech/machine-learnin…
With Mobius Labs' Aana models, we're bringing deeper multimodal understanding to Dropbox Dash

With Mobius Labs' Aana models, we're bringing deeper multimodal understanding to Dropbox Dash

Dropbox welcomes Mobius Labs to advance Dash’s multimodal AI, integrating Aana’s efficient architecture to enhance photo and video understanding at Dropbox scale.

dropbox.tech
2
7
2
23
43,342
Mobius Labs retweeted
mobicham (@mobicham) profile picture
mobicham@mobichamMay 15
github.com/dropbox/gemlite/t… You can now easily load various pre-quantized models (block FP8, NVFP4, AWQ/GPTQ/HQQ, certain GGUFs) via a vLLM plugin! You can also run on-the-fly quant as well, easy to use: 1 or 2 flags to enable!
1
4
19
1,995
Mobius Labs retweeted
mobicham (@mobicham) profile picture
mobicham@mobichamMay 12
Skimming through the PolarQuant paper by Google and found HQQ is still alive 👀
Image from mobicham (@mobicham)
1
6
22
1,873
Mobius Labs retweeted
mobicham (@mobicham) profile picture
mobicham@mobichamApr 24
Babe, wake up, new GemLite update. Up to 1.7x+ faster FP8 block quantization on the RTX PRO 6000 end-2-end in vLLM!
Image from mobicham (@mobicham)
1
3
32
2,470
Mobius Labs retweeted
Josh Clemm (@joshclemm) profile picture
Josh Clemm
@joshclemmApr 16
Open sourcing something fun from @Dropbox: Witchcraft. It's a local search engine built in Rust with no API keys or vector DB required. Think: ColBERT / late interaction style retrieval, but packaged to run locally (perfect for coding agents). Let's dive in👇
19
37
7
468
115,703
Mobius Labs retweeted
Pamela Fox (@pamelafox) profile picture
Pamela Fox
@pamelafoxApr 9
This was a great talk at DSpy meetup about how Dropbox optimized their search relevance judge: youtube.com/watch?v=gGGCJWbq… Also available in blog post form: dropbox.tech/machine-learnin… DSpy both improved structural conformance of SLM output (valid JSON) and agreement with humans (NMSE).
2
8
1
63
4,793
Wow! Quite a detailed blog from @Dropbox on how they optimize Dropbox Dash relevance judge with @DSPyOSS. Shout out to the Dropbox Dash team Eric Wang, Dmitriy Meyerzon, and @joshclemm !
Image from Omar Khattab (@lateinteraction)
Dropbox (@Dropbox) profile pictureDropbox
@DropboxMar 17
How we used DSPy to turn our relevance judge into a measurable optimization loop, making it more reliable and scalable in Dropbox Dash.

14
19
194
27,600
Mobius Labs retweeted
mobicham (@mobicham) profile picture
mobicham@mobichamFeb 16
New blogpost 🫡! dropbox.tech/machine-learnin…
How low-bit inference enables efficient AI

How low-bit inference enables efficient AI

Making products like Dropbox Dash accessible to individuals and businesses means tackling new challenges around efficiency and resource use.

dropbox.tech
1
4
9
794
Mobius Labs retweeted
mobicham (@mobicham) profile picture
mobicham@mobichamFeb 1
hmm isn't it obvious this would work better than QAT? I’ve been doing this already for HQQ+, the DeepSeek re-distill work, and FP4 weight quant quality recovery. More comments below 🧵
NVIDIA AI Developer (@NVIDIAAIDev) profile pictureNVIDIA AI Developer
@NVIDIAAIDevJan 28
We just launched an ultra-efficient NVFP4 precision version of Nemotron 3 Nano that delivers up to 4x higher throughput on Blackwell B200. Using our new Quantization Aware Distillation method, the NVFP4 version achieves up to 99.4% accuracy of BF16. Nemotron 3 Nano NVFP4: nvda.ws/4t63z9y Tech Report: nvda.ws/4bj3pp0
Image from NVIDIA AI Developer (@NVIDIAAIDev)
6
4
20
2,582
If you’re at @EurIPSConf this week and looking for your next ML role, ping us. Few of us from @Dropbox including @mobicham @appughar will be there Happy to chat about roles in multimodal, search, and video-AI.
Mobius Labs (@Mobius_Labs) profile pictureMobius Labs@Mobius_Labs21 Nov 2025
Replying to @Mobius_Labs
Looking for people with experience in : Building and scaling multimodal ML systems, Retrieval, ranking, and creative search problems, Turning prototypes into polished products 🌍 Remote (Germany or Denmark) Apply here: jobs.dropbox.com/listing/738… (Please use the link, not DMs!)
3
4
655
We are looking for a Machine Learning Engineer (IC3) to join our Multimedia AI team at Dropbox! You’ll work on the ML behind Dropbox Dash, building real multimodal search experiences across images, video, and text — and ship things that millions of people actually use.
1
253
Looking for people with experience in : Building and scaling multimodal ML systems, Retrieval, ranking, and creative search problems, Turning prototypes into polished products 🌍 Remote (Germany or Denmark) Apply here: jobs.dropbox.com/listing/738… (Please use the link, not DMs!)
1
802
Replying to @Dropbox
@Dropbox is acquiring our IP, and the Mobius core tech team is joining them. A huge milestone for us! dropbox.tech/machine-learnin…
With Mobius Labs' Aana models, we're bringing deeper multimodal understanding to Dropbox Dash

With Mobius Labs' Aana models, we're bringing deeper multimodal understanding to Dropbox Dash

Dropbox welcomes Mobius Labs to advance Dash’s multimodal AI, integrating Aana’s efficient architecture to enhance photo and video understanding at Dropbox scale.

dropbox.tech
2
7
2
23
43,342
And for my fellow nerds: with Dash, we're obsessed with depth of context, not just breadth. Dash isn't just calling APIs and doing naive keyword search. We build a full semantic index of all your content that is natively multimodal: not just documents but images, videos, everything. This is the only way to build true understanding and close the context gap to make AI actually useful at work.
Drew Houston (@drewhouston) profile pictureDrew Houston
@drewhouston23 Oct 2025
Our industry is spending trillions on AI models that can explain quantum physics but can't find your Q2 board deck. We're fixing this with Dropbox Dash, an AI assistant and search engine that knows about you, your work, and your company, and is deeply connected to all of your work apps.
2
4
34
40,225
Mobius Labs retweeted
Josh Clemm (@joshclemm) profile picture
Josh Clemm
@joshclemm23 Oct 2025
The best AI apps need the right context and @Dropbox Dash is built to provide it. Thrilled to welcome the incredible tech and talent from @Mobius_Labs to the team. Multimodal AI will make the Dash context platform even stronger 💪
Mobius Labs (@Mobius_Labs) profile pictureMobius Labs@Mobius_Labs23 Oct 2025
Replying to @Dropbox
@Dropbox is acquiring our IP, and the Mobius core tech team is joining them. A huge milestone for us! dropbox.tech/machine-learnin…
With Mobius Labs' Aana models, we're bringing deeper multimodal understanding to Dropbox Dash

With Mobius Labs' Aana models, we're bringing deeper multimodal understanding to Dropbox Dash

Dropbox welcomes Mobius Labs to advance Dash’s multimodal AI, integrating Aana’s efficient architecture to enhance photo and video understanding at Dropbox scale.

dropbox.tech
1
3
14
27,442
Mobius Labs retweeted
mobicham (@mobicham) profile picture
mobicham@mobicham20 Oct 2025
Looking forward to the @GPU_MODE hackathon later this Friday! Your boi will be there and happy to assist 🫡
Image from mobicham (@mobicham)
1
1
24
1,358
Mobius Labs retweeted
mobicham (@mobicham) profile picture
mobicham@mobicham16 Oct 2025
Kinda disappointed by Qwen3-VL's decoder perf with vLLM so far. Can squeeze better perf via GemLite instead of DeepGEMM but the numbers still seem suspiciously low
Image from mobicham (@mobicham)
1
4
604
Mobius Labs retweeted
mobicham (@mobicham) profile picture
mobicham@mobicham14 Sep 2025
It's happening: MXFP on sm_120 using pure Triton via GemLite, reaching very competitive performance 🫡 Shoutout to Hyunsung Lee (@HyunsungLe78860) for the hard work in bringing MXFP support to sm_120!
Image from mobicham (@mobicham)
2
10
53
3,339
Mobius Labs retweeted
mobicham (@mobicham) profile picture
mobicham@mobicham7 Aug 2025
Rewrote gpt-oss's experts forward pass (bfloat16 weights) in transformers to only use active experts + simple gemv MoE Triton kernel: 36 tokens/sec -> 165 tokens/sec (20B on the RTX6000 Pro) MXFP4 weights next, decoding should run faster 👀
3
10
75
3,368
Mobius Labs retweeted
mobicham (@mobicham) profile picture
mobicham@mobicham19 Jul 2025
Been playing with low-bit training this morning (3-bit, 2-bit, 1.58bit). Simple key trick to make it work: dumpen the gradients in the backward pass 🫡 gist.github.com/mobicham/b5c…
Low-bit training toy example

Low-bit training toy example

Low-bit training toy example. GitHub Gist: instantly share code, notes, and snippets.

gist.github.com
2
3
31
1,347
Mobius Labs retweeted
mobicham (@mobicham) profile picture
mobicham@mobicham7 Jul 2025
Simple and fast MXFP8 activation quant kernel: ✓ Padding-aware for arbitrary seq lens ✓ SM-aware unrolling to improve occupancy (large batches) ✓ ~10% faster than torch.compile in end-to-end inference
Image from mobicham (@mobicham)
1
11
92
4,725
Load more