Pinned Tweet
Dear users,
If you’d like to have early access to the Private LLM 2 Beta on iOS:
1. Update Private LLM 1 to the latest App Store release.
2. Settings > About.
3. Tap this icon.
For anyone who hasn't tried this out yet: Install the latest Private LLM 2 beta, call a friend and open a Speak conversation in the app. Preferably with one of the 12 character voices. 😅
None of our apps were affected. Because we don’t use any tracking SDKs or analytics in our apps. Privacy always comes first.
Like any distilled model, Ternary Bonsai 2 27B inherits refusals and censorship from its teacher model: Qwen 3.8 27B. Fortunately, there’s a very simple and elegant way to reduce or perhaps even completely eliminate them. 🧵
All it takes is a 5120-dimensional vector (the model’s hidden size) that is applied to every transformer layer’s activations.
In practice this is just a 20kb file that can uncensor/abliterate the Ternary Bonsai 2 27B model.
1
2
297
h/t @OrcaRouter for the refusal removal direction vector.
nitter.cf/OrcaRouter/status/2100…
PrismML team fit a 27B model into 5.9 GB.
We uncensored it without changing a single weight. 🐳
Introducing OrcaRouter Ternary Bonsai 2 27B Uncensored.
Traditional abliteration modifies weights. On a ~1.72-bit ternary model, that means re-quantization — potentially destroying the quality preserved by QAT.
So we moved abliteration into the runtime.
→ 0 weights modified
→ 0 re-quantization
→ original 5.9 GB pack stays bit-identical
→ 129 residual intervention sites
→ adjustable at inference
→ runs locally on Apple Silicon
No modified checkpoint. Bring the original Bonsai pack + our runtime.
Open source: github.com/Continuum-AI-Corp…
3
388
Some technical details on how we do this, without revealing too much to low-IQ copycats like @adrgrondin.
We losslessly repack the prism-ml/Ternary-Bonsai-2-27B-mlx-2bit model from 8 GiB to 7.6 GiB. On iPhones with 12 GB of RAM, MTLDevice.recommendedMaxWorkingSetSize is 8 GiB.
skill issue
Here’s Bonsai 2 27B running on an iPhone 17 Pro.
Textbook case of skill issue (bad quant and sampler). Same model with better mobile quants and a better sampler (Min-p).
1
7
917
Sandboxing is mandatory for all apps on the Mac App Store. We also enable the hardened runtime, even though that’s optional. These measures exist to protect you and your computer. Please think twice before downloading random macOS apps from Reddit, Gumroad etc.
🚨PSA: If you think you're a targeted individual, don't install macOS apps from the web. macOS code signing and TCC are broken. We accidentally found a bug that lets any command modify the binaries of other apps, including Signal, Brave, Chrome, and even Xcode. Watch the demo👇
This happens entirely in the background, with no password prompts or visible alerts. The modified binaries can then access the original app's keychain records and protected containers. We reported this to Apple two weeks ago and haven't received a response. We previously said we wouldn't contact Apple's bounty program again, but we believe it's necessary in this case to help protect Mac users. It affects macOS 26 and 27.
This bug doesn't impact apps downloaded from the Mac App Store. Many developers distribute apps on the web instead because Apple controls the App Store's review process and payment methods.
Private LLM retweeted
We don’t make BS farcical claims about 99.98% accuracy. We just build the best AI generated text detector that also happens to run on your iPhone’s Neural Engine.
Only a concern if you’re using unsandboxed apps like Ollama and LM Studio for local LLM inference. IYKYK. 💀
⚠️ Attackers poisoned Hugging Face & ClawHub (OpenClaw) with 575+ malicious skills from just 13 accounts.
🔸 Fake helpful AI tools that install trojans, miners & stealers (Windows + macOS)
🔸 Use hidden commands & indirect prompt injection
Quick action: Never install random AI skills or models. Always verify the source.
Read: thehackernews.com/2026/05/we…
1
740
Private LLM retweeted
Coming soon in the next update to Clean Links for macOS. Spot and clean tracking URLs on any website. Works fully offline. Also available on iOS.






























