ボブにゃんVRC retweeted
🚨 @ZAI_ORG JUST DROPPED GLM-5.2, AND IT IS PUNCHING RIGHT AT THE LEVEL OF CLAUDE OPUS 4.8 🤯
The kicker?
It’s a 753B parameter model with a true 1M-token context, released fully open-source under an MIT license
What makes this release technically interesting:
→ IndexShare Attention: They reuse a single indexer across every 4 sparse layers, cutting per-token FLOPs by 2.9× at a 1M context.
→ Better Speculative Decoding: An improved MTP layer increases acceptance length by up to 20%.
→ Adjustable Compute: Flexible thinking-effort levels let you explicitly trade off between performance and latency.
It’s also Day-0 ready.
You don’t have to wait for the ecosystem to catch up, it’s already supported in transformers, vLLM, and SGLang.
Repo, weights and paper in 🧵 ↓
2
2
11
2,269
Ille retweeted
Jun 15
Reporter: ‘Are you the one who makes decisions about Israel, or is it Trump and America?
Netanyahu: ‘I set certain parameters for our activity and we are doing it, but we cannot completely go against what the U.S. says’
1
2
4
191
Aibaogun Izeokhai O. retweeted
Jun 12
Why do Apple Intelligence models only support the iPhone 17 Pro?
I was interested how Apple managed to fit a 20B parameter model into the iPhone 17 Pro, a phone with only 12 gigs of RAM
Turns out they built a custom architecture only for Apple Silicon that allows only parts of the model to be loaded into memory. It's a twist on MoE, where they don't need the full model to be used!
According to Apple, a smaller model selects a fixed set of experts during initial processing. Experts that aren't needed aren't pulled into memory. "The 'routed experts' are swapped in when needed."
That's insane: they made it so that they don't even need the full model in use, but when they need other parts of it, they swap it out. Imagine a car that can swap out the engine while driving, depending on when it needs to be performant or efficient.
And since it's a multimodal model, that's what runs the new Siri voice and advanced dictation. That's why you can't get the new Siri voice on older iPhones: they don't have enough memory or bandwidth.
I thought that was pretty neat, and I wanted to share. You can read more at machinelearning.apple.com/re…
2
6
95
7,238
hbar1000 ༼ つ ◕_◕ ༽つ retweeted
Jun 10
In the latest weekly bonzo bytes newsletter from @bonzo_finance
One of the governance proposals is the DOVU Risk Parameters Update.
It suggests safely expanding liquidity capacity for $DOVU, while preserving robust liquidation margins, adjusting supply caps and borrow limits, as well as adjusting interest rates for the $DOVU asset.
It’s live right now, and expected to be implemented this week.
Head over to gov.bonzo.finance to register, read all about it, and share your feedback.
1
13
65
2,101
Huijin retweeted
Jun 6
Matrix Upgraded. Swarm Expanded 🦉
Owl AI Agent OS evolved. We’ve unlocked refined models and broader execution parameters
⬩ More Agents: Expanded options
⬩ New Battlefields: Fresh deployment vectors
⬩ Amplified Yield: Higher velocity returns optimized for volatility
Deploy: dapp.owlowl.ai
13
13
13
131
