Sabitlenmiş Tweet
In case you missed it:
Add a colon to your password, ":", because all the stealer logs have colons, so it'll end up splitting your password incorrectly.
English
Vincent Yiu
36.8K posts
Vincent Yiu
@vysecurity
Director, Red Team / Offensive Security. Help organizations safeguard their businesses from the bad guys.
Hong Kong Katılım Aralık 2014
357 Takip Edilen31.7K Takipçiler
@SkelSec @sekurlsa_pw I do rarely but never really understood why I needed it
English
Alright got the first Qwen3.7-Plus TPS wrong. Here's the real results.
English
Vincent Yiu retweetledi
Qwen3.6-35B-A3B-heretic-NVFP4 crushing 870 concurrent tok/s!!!
870 TOK/S!
ON A SINGLE DGX SPARK!
~115 Tok/s Single Stream
This is also under a grueling challenging benchmark and oddly it scored high on the most challenging GOD MODE category.
aeon-bench.com/share/aeon-7__…
English
@WescheNex1q @NVIDIAAI Sure:
🚀 vllm/nvidia/Qwen3.6-35B-A3B-NVFP4-MTP on ZEUS Bifrost — 32 parallel coding agents
• Total: 9.252 tokens in 11s
• Aggregate: 1267.3 tok/s peak · 879.9 tok/s sustained
• Per-stream: 49.6 high / 33.1 low / 42.5 avg tok/s
• TTFT 1.27s avg · E2E 9s avg
English
Vincent Yiu retweetledi
One DGX Spark. Qwen3.6-35B
64 users. 700+ tok/s. 32,768 tokens in 54 seconds.
38W
Each user has their own prompt and their own KV cache, and vLLM batches every active stream through the GPU each step.
Recipe → github.com/Weschera/spark… @NVIDIAAI
English
Maybe subsidise open source? That way more labs will release their models?
Jacob Wendler@jacob_wendler
Scoop: Talk has begun circulating around town that the White House may be considering a possible executive order on open-source AI, sparked by fears about Chinese dominance, nine people familiar told me, @BrendanBordelon, @delizanickel and @meredithllee
English
@compuficial Or… go out and meet people and get paid normal wages for normal work? 👏🙏
English
China is crazy. You could quit big tech, move into a dilapidated leaky Chengdu rooftop shack, train AI models as a freelancer, and still end up doing food delivery at night just to make rent.
English
That is some impressive crazy result for the Unsloth Qwen 3.6 35B NVFP4 dynamic - Fast .... it's a great throughput amplifer. it's be great to be part of the MoA Hermes setup..... a must have in your AI stack. @UnslothAI @NousResearch . these guys are pushing the consumer application to the max!!!!!!
mr-r0b0t@mr_r0b0t
Results 👀
English
Here's the speed you'd get on @NVIDIAAI DGX Spark running the new @UnslothAI Qwen3.6-35b-NVFP4-Fast in a single session 🔥
English
@UK_Daniel_Card I tried to make something like this 3 years ago or so, when ChatGPT first came out. Barely functioned. Your tool looks fantastic!
English
Bruh with that shitty vibe coded emoji? 🤣🤣🤣
The Hacker News@TheHackersNews
🚨 Hackers are using fake #Microsoft Entra passkey pages to target Microsoft 365 accounts. The vishing campaign steals credentials and MFA responses, then attempts to register attacker-controlled passkeys while adapting to prompts in real time. Read how the attack works: thehackernews.com/2026/07/hacker…
English
Or more like there’s no value in them being closed source. The only competitive angle of publishing a less than frontier model is by being open source or cheap.
Let’s see when China makes a model that’s more advanced than US frontier and it’s still OSS then I’ll eat my own words.
Jun Song@jun_song
Future of Open-Source AI Strategy: China pushed open-weight and aggressive low-pricing models in the past purely to close the tech gap. Now that GLM-5.2 took SoTA in performance-per-parameter, that gap has narrowed. According to reliable sources, China is now shifting toward AI export restrictions and closed weights. It makes sense—open-weight models just don’t make money and only burn cash. Nvidia is the only company with a real revenue initiative here, and they are fully committed to open source beyond just Nemotron. The only real counter-scenario to this is Huawei acquiring an open-source lab to push back. Other than that, we might see xAI drop small models to compete on price using their low-cost vertical full-stack, or Google release ultra-small models optimized for personal hardware. Personal edge device inference is still going to be a massive play.
English
Vincent Yiu retweetledi
