The fastest way to deploy mobile AI

Joined October 2025
Photos and videos
Cactus retweeted
Testing Cactus, a low-latency AI inference engine that runs a 1.2B model directly on an iPhone 12 Pro NPU using zero-copy memory mapping and a hybrid cloud router. See the full performance breakdown.
1
1
6
585
Cactus retweeted
Gemma 4 on Cactus can run realtime language, vision and speech on a mobile device...with multiple apps running in the background. Cactus routing algorithms gives Gemma the capacity to forward tasks to frontier models like Gemini and Claude when confused. Typical on-device AI demos are too structured, but this is probably the closest we've come to Jarvis-like AI assistant. @googlegemma @GeminiApp @cactuscompute
1
5
30
2,734
Proud to power on-device AI for Pebble โ€“ the product that pioneered the entire smart watch category โ€“ and now innovating on new wearable form factors.
Haven't posted much about @Pebble Index 01 recently (been...uh a bit in the weeds on PT2 shipping). But lots of progress has been made! Still need to order one? You can do that here ๐Ÿ˜‰ repebble.com/index We're in PVT2! Yes, 2 :) The first Production Verification Test (PVT) at end of March exposed two main problems 1) possible ESD damage during assembly was blowing out the BLE amp on some units, 2) a change we made to the mic waterproof membrane caused some units to fail audio testing. We fixed both with modifications to the assembly line. Things seem to be on a good track now with PVT2! When will my Index 01 ship? Our goal is to manufacture the first 2k units by mid-May and start shipping them out. It will take us roughly 3 months to manufacture and ship out all pre-orders, meaning we should be finished by July. As always, these are estimates. Delays may happen! Brushed Silver We are switching from offering a polished silver to brushed silver finish! It looks great - photo below, with lots more to come. Head to orders.rePebble.com to switch colours. Alpha testers Thanks to some brave souls who've been using early versions of Index 01 and reporting bugs, we've crushed a ton of software issues over the last 3 months! Performance and reliability are starting to be really smooth on iOS and Android. Software improvements All open source, of course! github.com/coredevices/mobilโ€ฆ - We added fully encrypted cloud backup (optional) - Local speech-to-text is working well, thanks to @cactuscompute parakeet-tdt-0.6b-v3! - You can optionally route audio or transcriptions to a webhook - enabling you to pipe recordings directly from your Index 01 to an agent like OpenClaw - Our lead engineer also added Home Assistant support (because she wanted it herself!!). - Added Beeper (text directly from Index 01!), music control (Android only), Notion, Apple Reminders and Google Tasks integration - Bring your own MCP server Still working on a lot of stuff - dramatically improved UI, more reminder app integrations (Todoist, Tasks.org), and more. Tell us - what are you excited to do with your Index 01? Do you have a favorite reminder/todo/notes app that you would like to use with it?
3
3
808
Cactus retweeted
Replying to @cactuscompute
@cactuscompute x @GoogleDeepMind x @ycombinator hosts the Gemma 4 Voice Agents Hackathon.
1
7
16
4,038
Cactus retweeted
The PowerSync AI Hackathon starts today. Bring your favorite AI ideas to life and compete for over $8k in prizes, including bonus prizes from our partners @supabase @neondatabase @mastra @tan_stack @cactuscompute. Let the hacking being!
1
1
7
666
Cactus retweeted
We now also have @cactuscompute on board as a partner in the hackathon! Cactus' on-device AI with cloud fallback pairs well with PowerSync. They are sponsoring a prize of a month of Cactus Hybrid inference for the best submission using Cactus ๐ŸŒต๐ŸŽ‰ x.com/powersync_/status/2026โ€ฆ

๐—”๐—ป๐—ป๐—ผ๐˜‚๐—ป๐—ฐ๐—ถ๐—ป๐—ด ๐˜๐—ต๐—ฒ ๐—ฃ๐—ผ๐˜„๐—ฒ๐—ฟ๐—ฆ๐˜†๐—ป๐—ฐ ๐—”๐—œ ๐—›๐—ฎ๐—ฐ๐—ธ๐—ฎ๐˜๐—ต๐—ผ๐—ป ๐Ÿค–๐Ÿง‘โ€๐Ÿ’ป We are hosting a virtual hackathon where the challenge is to build innovative AI-powered software using PowerSync as a sync engine. Join us to win up to $3,500 in cash prizes (and bonus partner prizes), explore cutting-edge AI use cases, connect with a vibrant community of builders, and discover new patterns for creating AI applications. More details coming soon, stay tuned! In the meantime, save the date and sign up for updates โฌ‡๏ธ
2
2
2,346
Cactus retweeted
From generalist to expert! See how @cactuscompute used #Tunix for Supervised Fine-Tuning on the Gemma 3 1B model, boosting its tool-calling capabilities from 28% to 35%. All on the free tier of Google Colab. #AI #SFT #Gemma goo.gle/gemma3-tunix-tuning
2
30
6,343
Cactus retweeted
Excited to share that our team placed 2nd at the Cactus (YC S25) x Nothing x Hugging Face Mobile AI Hackathon. We were up against teams from MIT, Stanford, and builders from around the world. Grateful to have had the chance to build and compete alongside so many talented people.
1
3
425
Cactus retweeted
4 Dec 2025
Spent 24 hours at Nothingโ€™s new London HQ as part of the @cactuscompute ร— @huggingface ร— @nothing hackathon. My team and I built Pulse; a first-aid assistant to guide bystanders through basic life support whilst reporting to first responders providing situational context. Against ~700 international participants, we were awarded the honorary prize!!!
1
1
581
Cactus retweeted
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 Grandmaโ€™s phones will run them. Stay tuned! github.com/cactus-compute/caโ€ฆ
7
13
156
37,461
Cactus retweeted
26 Nov 2025
We just shipped the Cactus React Native SDK๐ŸŒต- the fastest and most efficient on-device AI inference engine for React Native.โšก๏ธLightweight, insanely fast, and built for mobile devices from the ground up.๐Ÿš€
1
2
3
575
Cactus retweeted
Hackathon alert! London, SF, Boston. This Friday! ๐Ÿ‘€ @nothing is teaming up with @cactuscompute and @huggingface to hack on redefining on-device AI experiences! Come build something memorable, meet the teams, and ship in 24 hours! Signups are wild so far ๐Ÿ”ฅ
9
20
200
48,435
21 Nov 2025
Cactus React Native v1 is live! Deploy AI on-device with text inference, tool calling, embeddings and more โ€“ powered by the fastest edge inference engine ๐ŸŒต Our React Native bindings run on @margelo_com's Nitro Modules, yielding the fastest mobile inference we've seen so far.
1
2
3
379