Multimodal AI for the world's scale. Proponents of Open Source and Open Intelligence. https://xitter.twitterapi.xyz/t.co/1nC6r8hOrE for some of our recent work.
Berlin, Germany
Joined April 2018
- Tweets420
- Following214
- Followers3.3K
- Likes1.5K
Pinned Tweet
@Dropbox is acquiring our IP, and the Mobius core tech team is joining them. A huge milestone for us!
dropbox.tech/machine-learnin…
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.
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!)
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
254
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!)
FP4 weights meets high accuracy: Logit‐distillation bias correction for MXFP4 & NVFP4. On Llama-3.1-8B recovers ≥99% relative quality. Details at: mobiusml.github.io/fp4_blogp…
GemLite now supports @vllm_project vllm V1, which brings up to 1.25x faster inference speed vs V0!
github.com/mobiusml/gemlite/…
Initial attempts on brining Gemlite to @AMD MI300X
Hitting ~200 tokens/sec with Llama3-8B (bs=1) using GemLite (after changes) on the AMD MI300X.
Performance lands between the A100 and H100 — given the price point, the MI300X is shaping up to be a very compelling inference option!
5
295
We just added bfloat16 support to GemLite! Took a few technical hops to get atomic addition working smoothly in Triton—details here: mobiusml.github.io/gemlite_b…
Thanks to bfloat16 support, GemLite now runs models like Google’s Gemma3—avoiding the usual accuracy loss from fp16 conversions. huggingface.co/mobiuslabsgmb…
6
262
New releases for hqq ( github.com/mobiusml/hqq/rele……) and gemlite ( github.com/mobiusml/gemlite/…… )!
The most exciting update is that we're bringing support to @vllm_project
You can quantize any model on the fly in @vllm_project using HQQ. It should work with most models, including vision-language models.
338
We’ve improved DeepSeek R1 distilled models using logits distillation—delivering +4-14% gains on GSM8K while only spending $3-18 per training run! 🚀
Now available on Hugging Face-e—run them efficiently with HQQ! 🤗
🔗More details at: mobiusml.github.io/r1_redist…
Models 🔥
✅New: Llama3-8B Re-Distilled (+14% GSM8K, better reasoning!) huggingface.co/mobiuslabsgmb…
✅Qwen1.5B & Qwen7B Re-Distilled (released earlier)
3
358
We are releasing more re-distilled R1 models. Now including a Qwen-7B with a GSM8K score of 83.4%, +4.55% higher than that of the original distilled model.
7B: huggingface.co/mobiuslabsgmb…
1.5B: huggingface.co/mobiuslabsgmb…
Our re-distilled @deepseek_ai R1 (1.5B) outperforms the original distilled model! Get it at huggingface.co/mobiuslabsgmb…. We’re distilling more models and look forward to releasing them soon!
It was a real pleasure and a productive experience collaborating with the team at
@PyTorch and the SGlang team at
@lmsysorg to bring gemlite to TorchAO and SGLang. This is an exciting area that requires a lot of work. Let’s push the boundaries of efficient computing!
Scaling Large Language Models (#LLMs) is tough due to resource demands. Quantization helps but often comes with trade-offs in accuracy and performance.
Discover how #PyTorch, Mobius Labs, and SGLang tackled these challenges with Gemlite, TorchAO, and SGLang for efficient low-precision inference.
📖 Read here: hubs.la/Q033z_kw0
github.com/mobiusml/gemlite is a set of Triton Kernels for Efficient Low-Bit Matrix Multiplication and meant to easy to understand, while giving high performance. If efficient inference is your area of interest, we are eager to hear more!
197
Releasing a new version of Gemlite github.com/mobiusml/gemlite significantly improved performance on datacenter GPUS (A100/H100) delivering up to 7–8x faster prefill and 3–6x faster decoding compared to PyTorch's tinygemm.
Introduced new kernels, max-autotuning, and several other improvements Check out the release details at github.com/mobiusml/gemlite/…
1
1
170
Really happy to contribute to the batched version of faster-whisper that is 4x faster and more accurate 🚀🚀🚀 github.com/SYSTRAN/faster-wh…
Shout out to @SebastianJilt and also github.com/MahmoudAshraf97
A bit more details at github.com/SYSTRAN/faster-wh…
1
166
Check out our own @mobicham's lecture on building low-bit Triton kernels & the impressive speedups we’re achieving with them, presented at @GPU_MODE channel. Also includes plenty of tips & tricks for performance tuning in Triton & GPU programming.
youtube.com/watch?v=7c3c3bCG…
@mobicham will be talking about low-bit Triton kernels and his experience building github.com/mobiusml/gemlite this Saturday at the GPU mode discord. It’s sure to be a great session if you’re interested in efficiency!
discord.gg/KKkmtQPx?event=12…
If anyone is interested in using Whisper-Turbo with Faster-Whisper, you can use this model from our Hugging Face hub: huggingface.co/mobiuslabsgmb…
