https://xitter.cf/t.co/YJ88R0IWCp | On-device foundation models. High intelligence, low memory, low power

San Francisco Bay Area
Joined September 2026
๐—˜๐—ฑ๐—ด๐—ฒ๐Ÿฌ ๐—”๐—œ ๐—ฟ๐—ฎ๐—ป๐—ธ๐˜€ #๐Ÿฐ๐Ÿฑ ๐—ฎ๐—บ๐—ผ๐—ป๐—ด ๐˜๐—ต๐—ฒ ๐—ณ๐—ฎ๐˜€๐˜๐—ฒ๐˜€๐˜-๐—ด๐—ฟ๐—ผ๐˜„๐—ถ๐—ป๐—ด ๐—ฒ๐—บ๐—ฒ๐—ฟ๐—ด๐—ถ๐—ป๐—ด ๐—ผ๐—ฝ๐—ฒ๐—ป-๐˜€๐—ผ๐˜‚๐—ฟ๐—ฐ๐—ฒ ๐—ผ๐—ฟ๐—ด๐—ฎ๐—ป๐—ถ๐˜‡๐—ฎ๐˜๐—ถ๐—ผ๐—ป๐˜€ ๐—ผ๐—ป ๐—š๐—ถ๐˜๐—›๐˜‚๐—ฏ! Put together by @commit and @Supabase, we're featured in the OSSCAR Q3 2026 Emerging ranking. A big thank you to everyone who's explored our projects, starred our repos, or shared feedback. It's great to see our work being seen in the open-source community. And to all the other teams on the list: hello! ๐Ÿ‘‹ We're always up for conversations about efficient inference, on-device AI, and what it takes to make powerful models run on everyday hardware.
1
1
7
106
๐—ฆ๐—ฒ๐—ฒ ๐—น๐—ถ๐˜€๐˜: osscar.dev/org/edge0-ai ๐—ง๐—ฟ๐˜† ๐—˜๐—ฑ๐—ด๐—ฒ๐Ÿฌ ๐—ผ๐—ป ๐—š๐—ถ๐˜๐—›๐˜‚๐—ฏ: github.com/Edge0-AI ๐—๐—ผ๐—ถ๐—ป ๐—ผ๐˜‚๐—ฟ ๐——๐—ถ๐˜€๐—ฐ๐—ผ๐—ฟ๐—ฑ to get help with setup or share what you're building: discord.gg/Tg3W2FEuQ
2
19
๐—” ๐Ÿ‘๐Ÿ“๐—•-๐—ฝ๐—ฎ๐—ฟ๐—ฎ๐—บ๐—ฒ๐˜๐—ฒ๐—ฟ ๐—บ๐—ผ๐—ฑ๐—ฒ๐—น, ๐—ฟ๐˜‚๐—ป๐—ป๐—ถ๐—ป๐—ด ๐—น๐—ผ๐—ฐ๐—ฎ๐—น๐—น๐˜† ๐—ผ๐—ป ๐—ฎ๐—ป ๐—ถ๐—ฃ๐—ต๐—ผ๐—ป๐—ฒ ๐Ÿ๐Ÿ” ๐—ฃ๐—ฟ๐—ผ: ๐Ÿ”.๐Ÿ’ ๐˜๐—ผ๐—ธ/๐˜€ ๐—ฑ๐—ฒ๐—ฐ๐—ผ๐—ฑ๐—ฒ. ๐Ÿ“ฑ The obstacle: weight memory. 35B parameters (weights) ร— 16 bits = ~70 GB. ๐—ฆ๐˜๐—ฒ๐—ฝ ๐Ÿ ๐—ผ๐—ณ ๐—ณ๐—ถ๐˜๐˜๐—ถ๐—ป๐—ด ๐—ถ๐˜ ๐—ผ๐—ป ๐—ฎ ๐—ฝ๐—ต๐—ผ๐—ป๐—ฒ: ๐˜€๐—ต๐—ฟ๐—ถ๐—ป๐—ธ ๐˜๐—ต๐—ฒ ๐˜„๐—ฒ๐—ถ๐—ด๐—ต๐˜๐˜€. ๐Ÿงต (We made it work and optimized it for iPhone 16 Pro, iOS 26.6.2.) 1/5
1
1
7
164
Step 1 done: less to store and load. But RAM is untouched. Sparsity cuts compute per token, not the bytes you have to hold, and ~19.5 GB is far more than a phoneโ€™s RAM. The next step is running the model without keeping that entire checkpoint in RAM. Stay tuned for our next post on how we do it (hint: MoE) 5/5
1
16
๐—ง๐—ฟ๐˜† ๐—˜๐—ฑ๐—ด๐—ฒ๐Ÿฌ ๐—ผ๐—ป ๐—š๐—ถ๐˜๐—›๐˜‚๐—ฏ: github.com/Edge0-AI ๐—๐—ผ๐—ถ๐—ป ๐—ผ๐˜‚๐—ฟ ๐——๐—ถ๐˜€๐—ฐ๐—ผ๐—ฟ๐—ฑ to get help with setup or share what youโ€™re building: discord.gg/Tg3W2FEuQ
17
Think your laptop's limited RAM is a barrier to your AI dreams? Think again! ๐—›๐—ฒ๐—ฟ๐—ฒ ๐—ฎ๐—ฟ๐—ฒ ๐Ÿ— ๐—ถ๐—ป๐—ฐ๐—ฟ๐—ฒ๐—ฑ๐—ถ๐—ฏ๐—น๐—ฒ ๐—ผ๐—ฝ๐—ฒ๐—ป ๐—”๐—œ ๐—บ๐—ผ๐—ฑ๐—ฒ๐—น๐˜€ ๐˜†๐—ผ๐˜‚ ๐—ฐ๐—ฎ๐—ป ๐—ฒ๐˜…๐—ฝ๐—น๐—ผ๐—ฟ๐—ฒ, ๐—ฒ๐˜ƒ๐—ฒ๐—ป ๐˜„๐—ถ๐˜๐—ต ๐—บ๐—ถ๐—ป๐—ถ๐—บ๐—ฎ๐—น ๐—ฟ๐—ฒ๐˜€๐—ผ๐˜‚๐—ฟ๐—ฐ๐—ฒ๐˜€, ๐—ฎ๐˜€ ๐˜„๐—ฒ ๐˜€๐˜๐—ฒ๐—ฝ ๐—ถ๐—ป๐˜๐—ผ ๐—ข๐—ฐ๐˜๐—ผ๐—ฏ๐—ฒ๐—ฟ ๐Ÿ๐ŸŽ๐Ÿ๐Ÿ”. ๐—Ÿ๐—ฒ๐˜โ€™๐˜€ ๐—ฑ๐—ถ๐˜ƒ๐—ฒ ๐—ถ๐—ป! ๐Ÿงต๐Ÿ‘‡
2
3
18
147
๐Ÿ There's no universal winner. RAM, latency, language coverage, and licenses all affect the choice. ๐—ฃ๐—ถ๐—ฐ๐—ธ ๐˜„๐—ต๐—ฎ๐˜ ๐—ณ๐—ถ๐˜๐˜€ ๐˜†๐—ผ๐˜‚๐—ฟ ๐˜„๐—ผ๐—ฟ๐—ธ๐—น๐—ผ๐—ฎ๐—ฑ ๐—ฎ๐—ป๐—ฑ ๐—ต๐—ฎ๐—ฟ๐—ฑ๐˜„๐—ฎ๐—ฟ๐—ฒ. ๐—ช๐—ต๐—ฎ๐˜ ๐˜„๐—ผ๐˜‚๐—น๐—ฑ ๐˜†๐—ผ๐˜‚ ๐—ฎ๐—ฑ๐—ฑ? ๐˜Š๐˜ถ๐˜ณ๐˜ข๐˜ต๐˜ฆ๐˜ฅ ๐˜ฃ๐˜บ ๐˜Œ๐˜ฅ๐˜จ๐˜ฆ0. ๐˜•๐˜ฐ๐˜ต ๐˜ข ๐˜ณ๐˜ข๐˜ฏ๐˜ฌ๐˜ช๐˜ฏ๐˜จ.
1
1
18
๐—ง๐—ฟ๐˜† ๐—˜๐—ฑ๐—ด๐—ฒ๐Ÿฌ ๐—ผ๐—ป ๐—š๐—ถ๐˜๐—›๐˜‚๐—ฏ: github.com/Edge0-AI ๐—๐—ผ๐—ถ๐—ป ๐—ผ๐˜‚๐—ฟ ๐——๐—ถ๐˜€๐—ฐ๐—ผ๐—ฟ๐—ฑ to get help with setup or share what youโ€™re building: discord.gg/Tg3W2FEuQ
1
34
๐—ง๐—ฟ๐˜† ๐—˜๐—ฑ๐—ด๐—ฒ๐Ÿฌ ๐—ผ๐—ป ๐—š๐—ถ๐˜๐—›๐˜‚๐—ฏ: github.com/Edge0-AI ๐—๐—ผ๐—ถ๐—ป ๐—ผ๐˜‚๐—ฟ ๐——๐—ถ๐˜€๐—ฐ๐—ผ๐—ฟ๐—ฑ to get help with setup or share what youโ€™re building: discord.gg/Tg3W2FEuQ
15
๐Ÿ There's no universal winner. RAM, latency, language coverage, and licenses all affect the choice. ๐—ฃ๐—ถ๐—ฐ๐—ธ ๐˜„๐—ต๐—ฎ๐˜ ๐—ณ๐—ถ๐˜๐˜€ ๐˜†๐—ผ๐˜‚๐—ฟ ๐˜„๐—ผ๐—ฟ๐—ธ๐—น๐—ผ๐—ฎ๐—ฑ ๐—ฎ๐—ป๐—ฑ ๐—ต๐—ฎ๐—ฟ๐—ฑ๐˜„๐—ฎ๐—ฟ๐—ฒ. ๐—ช๐—ต๐—ฎ๐˜ ๐˜„๐—ผ๐˜‚๐—น๐—ฑ ๐˜†๐—ผ๐˜‚ ๐—ฎ๐—ฑ๐—ฑ? ๐˜Š๐˜ถ๐˜ณ๐˜ข๐˜ต๐˜ฆ๐˜ฅ ๐˜ฃ๐˜บ ๐˜Œ๐˜ฅ๐˜จ๐˜ฆ0. ๐˜•๐˜ฐ๐˜ต ๐˜ข ๐˜ณ๐˜ข๐˜ฏ๐˜ฌ๐˜ช๐˜ฏ๐˜จ.
2
๐Ÿ”Š ๐—ง๐—ฒ๐˜…๐˜-๐˜๐—ผ-๐—ฆ๐—ฝ๐—ฒ๐—ฒ๐—ฐ๐—ต (๐—ง๐—ง๐—ฆ) โ€ข ๐—”๐˜‚๐—ฑ๐—ถ๐—ผ๐Ÿ–-๐—ง๐—ง๐—ฆ ๐ŸŽ.๐Ÿ”๐—• โ€” 11 languages + zero-shot voice cloning huggingface.co/Edge0/Audio8-โ€ฆ โ€ข Chatterbox V3 โ€” expressive multilingual TTS huggingface.co/ResembleAI/chโ€ฆ โ€ข CosyVoice 3 โ€” streaming speech synthesis huggingface.co/FunAudioLLM/Fโ€ฆ
22
๐ŸŽ™๏ธ ๐—ฆ๐—ฝ๐—ฒ๐—ฒ๐—ฐ๐—ต ๐—ฅ๐—ฒ๐—ฐ๐—ผ๐—ด๐—ป๐—ถ๐˜๐—ถ๐—ผ๐—ป (๐—”๐—ฆ๐—ฅ) โ€ข ๐—”๐˜‚๐—ฑ๐—ถ๐—ผ๐Ÿด-๐—”๐—ฆ๐—ฅ-๐—œ๐—ป๐—ณ๐—ถ๐—ป๐—ถ๐˜๐—ฒ โ€” continuous streaming, bounded KV cache huggingface.co/Edge0/Audio8-โ€ฆ โ€ข X2Streaming-ASR โ€” append-only transcripts github.com/X-Square-Robot/X2โ€ฆ โ€ข Qwen3-ASR-0.6B โ€” multilingual ASR huggingface.co/Qwen/Qwen3-ASโ€ฆ
6
๐Ÿง  ๐—Ÿ๐—ผ๐—ฐ๐—ฎ๐—น ๐—Ÿ๐—Ÿ๐— ๐˜€ โ€ข ๐—˜๐—ฑ๐—ด๐—ฒ๐ŸŽ-๐Ÿ‘๐Ÿ“๐—• โ€” SSD-streamed experts to ease RAM pressure huggingface.co/Edge0/Edge0-3โ€ฆ โ€ข LFM2.5-24B-A2B โ€” 2B active parameters huggingface.co/LiquidAI/LFM2โ€ฆ โ€ข Qwen3.5-2B โ€” compact multimodal AI huggingface.co/Qwen/Qwen3.5-โ€ฆ
25
๐—ง๐—ฟ๐˜† ๐—˜๐—ฑ๐—ด๐—ฒ๐Ÿฌ ๐—ผ๐—ป ๐—š๐—ถ๐˜๐—›๐˜‚๐—ฏ: github.com/Edge0-AI ๐—๐—ผ๐—ถ๐—ป ๐—ผ๐˜‚๐—ฟ ๐——๐—ถ๐˜€๐—ฐ๐—ผ๐—ฟ๐—ฑ to get help with setup or share what youโ€™re building: discord.gg/Tg3W2FEuQ
1
25