Doughboy 💸

LFM2:3B in space, on Cluster Gate2: ✨ “This image is a highly detailed, close-up view of Earth as seen from space, likely captured by a satellite or space telescope. The Earth is depicted as a large, circular sphere with a predominantly blue hue, indicating the vast oceans that cover most of its surface. The blue is interspersed with swirling white clouds, which are particularly prominent over the landmasses, suggesting the presence of weather systems and atmospheric activity. The overall composition of the image highlights the beauty and complexity of our planet, showcasing the dynamic interplay between the oceans, atmosphere, and landmasses." Congratulations to @DPhiSpace for this incredible milestone! 🌎

Proud to partner with @MercedesBenz in a multi-year agreement to bring embedded, on-device intelligence to Mercedes-Benz vehicles, first in North America. This marks an important step toward making in-car AI more capable, more responsive, and more useful in everyday driving. At @liquidai, we believe the future of intelligence in the physical world depends on models that are fast, private, efficient, and able to run directly on the hardware already inside the system. In the vehicle, that means advancing speech, language understanding, and reasoning enabling more natural and robust conversational experiences for drivers and passengers. The software-defined vehicle is one of the most consequential real-world deployments of AI, and Mercedes-Benz has approached it with exactly the rigor this challenge deserves. Proud of what our teams are building together, and excited for the road ahead as we work toward production deployment in the second half of 2026. liquid.ai/press/liquid-a…

LFMs powering speech, language understanding, and reasoning directly inside the vehicle Very exciting times for edge models!

In the US, the average person spends >5 YEARS (!) of their lifetime sitting behind the wheel of a vehicle. If you include time spent as a passenger, that estimate easily doubles. I'm very proud of this partnership with @MercedesBenz — together we will be making in-car AI more capable, more responsive, and more useful to everyone. A truly AI-native vehicle is the perfect embodiment capturing the worldwide impact of massively multimodal on-device AI.

We’re entering a multi-year partnership with @MercedesBenz to scale embedded, on-device intelligence for their third- and fourth-generation MBUX. Our goal: to make the driver/vehicle relationship even more natural and effortless. Read more about our partnership: liquid.ai/press/liquid-a…

Today, we release LFM2.5-VL-450M, a vision-language model built for real-time reasoning on edge devices. It processes a 512×512 image and returns structured outputs in ~240ms on-device.

@boyuan_chen @neural_avb Fine-tuning them over well defined domains w/ relatively small amount of samples (order ~10k) leads to great results. At that parameter count the models need hand holding, but after that they are really great 🚀

@neural_avb 1.6B params in the browser on WebGPU is wild. At that size you can embed vision in literally anything - browser extensions, mobile apps, edge devices. Quality is the big question though. Anyone actually tested these on something beyond demos?

FYI these batshit fast VLMs are all open and available on huggingface. 1.6B params. You can run in a browser on webgpu using the onnx version. If you are on Mac, mlx-community has a version too. They even have a notebook with SFT and DPO recipes. 🙏🏼 huggingface.co/LiquidAI/LFM2.…

@paularambles Jan 2025, got my place for a bag of chips

moving to sf and signing a lease in october 2024 was my version of buying bitcoin in 2012

Today, we release LFM2.5, our most capable family of tiny on-device foundation models. It’s built to power reliable on-device agentic applications: higher quality, lower latency, and broader modality support in the ~1B parameter class. > LFM2.5 builds on our LFM2 device-optimized hybrid architecture > Pretraining scaled from 10T → 28T tokens > Expanded reinforcement learning post-training > Higher ceilings for instruction following 🧵

@michelleefang merry Christmas 🥹

if you're vibe coding or building over the holidays, i want to gift one of you a 6 month subscription of claude pro to support <3 just drop a comment below. merry christmas!

@Henry_Ndubuaku Very cool, wondering if you've benchmarked this against llamacpp. Are there perf gains vs ggml?

1.6B INT8 VLM by @liquidai on Cactus (YC S25) never exceeds 231MB of peak memory usage at any context size. 1. Cactus is aggressively optimised to run on budget devices with minimal resources, enabling efficiency, negligible pressure on your phone and passes your OS safety mechanisms. 2. Notice how 1.6B INT8 CPU reaches 95 toks/sec on Apple M4 Pro, faster than your eyes could process. Our INT4 will almost 2x the speed when merged. Expect up to 180 toks/sec decode speed. 3. The prefill speed reaches 513 toks/sec. Our NPU kernels will 5-11x that once merged. Expect up to 2500 - 5500 toks/sec. The time to first token of your large context prompt will take less than 1sec. 4. LFM2-1.2B-INT8 in the Cactus compressed format takes only 722mb. This means that with INT4 will shrink to 350mb. Almost half as much as GGUF, ONNX, Executorch, LiteRT etc. 5. Once done, we will start recommending 1B models to our users, cause your Grandma’s phones will run them. Stay tuned! github.com/cactus-compute…

@rohanpaul_ai @Meta Hi @danielhanchen :) can we finetune this on unsloth?

🧮 @Meta released MobileLLM-R1 for edge reasoning A sub-1B family that claims 2x to 5x gains on math and code versus other open models while staying small enough for constrained devices. The lineup spans 140M to 950M parameters and targets math, coding, and scientific reasoning rather than broad chat. On MATH500 it reports about 5x the accuracy of OLMo-1.24B and about 2x over SmolLM2-1.7B, and on GSM8K, AIME, and LiveCodeBench it matches or beats Qwen3-0.6B despite far fewer tokens. Context length is 4K for base models and 32K after post-training, which helps long problems but raises KV-cache memory needs when tokens pile up. Training uses about 4.2T tokens compared with Qwen3-0.6B at 36T, which is roughly 11.7% of the data to reach similar or better accuracy. For on-device runs, the small footprint and 4K context help latency and memory, while 32K mode likely needs careful quantization and KV-cache tuning.

@creatine_cycle guy maxx

your career goal should be to get to the point where you have "a guy" for stuff. nothing is more alpha than recommending a guy. tax? i know a guy legal? i have a guy for that 409 valuation? a guy agentic workflows? believe it or not there's a guy for that

@teodorio touching grass, the occasional bender, getting a gf

Ok so working 12 hour days and 4-5 on weekends with high context switching across really technical fields is unsustainable. I feel like my body and mental health is breaking down. Quick tips to keep it up?

@neural_avb I've often thought about asking this as an interview question. Surprising how many don't know where there HF models live

I have 90GB more disk space because I ran: rm ~/.cache/huggingface/* 😇

Introducing LFM2-VL — our new generation of efficient vision-language models for real-world deployment, from smartphones and laptops to wearables and embedded systems. 🧵

meet LFM2-VL: an efficient Liquid vision-language model for the device class. open weights, 440M & 1.6B, up to 2× faster on GPU with competitive accuracy, Native 512×512, smart patching for big images. efficiency is our product @liquidai download them on @huggingface: huggingface.co/LiquidAI/LFM2-… huggingface.co/LiquidAI/LFM2-… read the blog post: liquid.ai/blog/lfm2-vl-e…

@HPouransari any plans to port to other backends (onnx, llamacpp etc) ?

🌟Explore key insights from the FastVLM project (real-time vision-language model) in this blog post: machinelearning.apple.com/research/fast-…

@OfficialLoganK are log probs available via Gemini dev api?

@OfficialLoganK RPD for the 2.5 models is quite limited in Tier 1/2 still. Any plans to unleash this soon? 🥵

Gemini 2.5 Pro is back in the free tier of the API, have a great weekend : )