Free, open source, native Swift. No account needed to run local models.
github.com/osaurus-ai/osauru…
We built the investigative journalist agent to test one thing: can a tiny model do real work on sensitive files?
LFM3 did it on day 0, with the Wi-Fi off.
tpae retweeted
Journalists can't upload a leak to the cloud.
So we turned off the Wi-Fi.
A 1.7B model read the whole case file, followed the money, and cited its sources.
LFM3. Day 0 on Osaurus.
Introducing 𝙻𝙵𝙼𝟹: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱: text, 256k context window.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱: vision-language model with video-input support for spatial understanding
> 𝙻𝙵𝙼𝟹-𝙰𝚞𝚍𝚒𝚘-𝟸𝙱: tool calls, voice chat, reasoning in 18 languages.
> 𝙻𝙵𝙼𝟹-𝙾𝚖𝚗𝚒-𝟸𝙱: an omni model combines vision, audio, and language.
> 𝚍𝟷-𝟸𝙱 and 𝚍𝟷-𝙾𝚖𝚗𝚒-𝟸𝙱: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱 is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 is #1 on the Pareto frontier of device intelligence 𝚙𝚒𝚙𝚎𝚝𝚝𝚎.𝚕𝚒𝚚𝚞𝚒𝚍.𝚊𝚒
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ 🧵
Having worked with smart contracts for years, I keep coming back to one thing.
The biggest barrier to adoption was never the tech. It was the learning curve.
Agents might finally solve that. They can make informed decisions on the user's behalf and navigate the space for them.
I call my agent more than I call my friends now.
Huge work by @RaajeevCn on this one.
tpae retweeted
more ppl needa be talking about how it kills the whole point of being local if you have to go point claude to an engine to startup the model, and then get claude to once again wire the api and params into the config to even get it working? isnt this super redundant to the whole local ai thing? being local doesnt just mean running on your hardware but also being accessible for as many users as possible, do ppl really want to use tab indents and all spacing/syntax to write out a proper toml/json config file for usage in a harness
4. run ai locally and pay for electricity
3 ways enterprise ai plays out imo
1. models get good enough to justify the spend (muse, grok bot, starting in small biz where context needs are low)
2. big enterprises posttrain on proprietary data and keep hill climbing
3. model agnostic players with rich data access beat 1/2 (i expect consolidation)
onboarded a new customer today.
he's an accountant. doesn't trust any cloud with his clients' data. he's upgrading his machine just to run local models.
then he told me why: his last firm had to shut down after a security breach.
he's more paranoid than ever. honestly, he should be.
tpae retweeted
Nanbeige actually lived up to its expectations, with reasoning on it tended to over reason (not in a way of thinking about things that didn't matter too much like the Qwen series, but much more in a "plans out all the specifics even in a simple job" kind of way), and while I had it turned off for the demo, I know just from the tests I did that it lives up to the claims and competes with Qwen 3.8 27b as a heavy coding capable model.
While cache may be larger due to older arch, Osaurus' efficient cache handling allows for smooth turn by turn usage without your RAM usage going higher than the model's ~20gb.
tpae retweeted
Local-first is antifragile by default. When agents run on your own hardware, there’s no central endpoint to breach and no single outage that takes everyone down.
We need an "Antifragile Internet" built for the agentic era.
The hacks over the past few days (PHP, Railway, eMudhra, LiteLLM etc) show how fragile the existing Internet has become. AI agents are challenging the Internet's structure in 3 distinct ways:
1/ On the offense side, agents are a weapon - AI finds and exploits bugs on a scale not seen before, and the cost of attack asymptotes to 0. Bugs that used to take human experts weeks and months are easily found, adapted, and exploited by AI agents. Decades of legacy code that ran on "security by obscurity" no longer holds against agents.
2/ On the defense side, agents are a target - AI generated code is often bug-prone. AI agents themselves, as machines of unconstrained output are liable to prompt injections and other types of attacks that can leak keys, PII, and other pieces of information.
3/ On the identity side, agents are dominating Internet activity - this year, machine traffic has already passed human traffic, and may continue to grow 1000x over the next few years. Content and webpages used to human requests and behavior are breaking with agents, and it is increasingly difficult to tell human content from AI content.
All this means we will need to see a ground-up redesigning of the Internet, one with more cryptographic verifiability, identity protocols, rule-enforcement, and other security primitives that eliminate bug categories rather than simply patch them - an "Antifragile Internet" that gets stronger and more secure as more agents participate in it.
This is an emerging thesis area for us at @usv, if you're building in the space please reach out!
More thoughts soon.
