Andrew Zhu (@xhinker) header
Andrew Zhu (@xhinker) profile picture

on Personal AI | open source contributor | startup founder member | ex-Microsoft blog: https://teapawt.net/t.co/QjnjnNVoCD

WA, US
Joined August 2009
Photo by Andrew Zhu (@xhinker)Photo by Andrew Zhu (@xhinker)Photo by Andrew Zhu (@xhinker)Photo by Andrew Zhu (@xhinker)Photo by Andrew Zhu (@xhinker)Photo by Andrew Zhu (@xhinker)Photo by Andrew Zhu (@xhinker)Photo by Andrew Zhu (@xhinker)Photo by Andrew Zhu (@xhinker)
Andrew Zhu retweeted
🤖 Made with AI
84
879
68
9,949
373,943
Andrew Zhu retweeted
Here’s why they crucified Jensen btw Don’t let them psyop you, they have nothing special.
Back to the Permanent Underclass we go!
88
492
34
6,495
576,615
Andrew Zhu retweeted
The mission has not and will not change
92
344
25
2,456
69,224
Andrew Zhu retweeted
The exposes on Jensen Huang are part of a long-running campaign by the Democrats/Anthropic to harm open source and give Anthropic an AI monopoly heavily regulated for the political left. Suffocating nanny surveillance state AI is the final goal. Censorship regime on steroids.
74
353
17
1,979
53,021
Andrew Zhu retweeted
Unsloth Dynamic V3.1 (preview) UD-IQ4_XS VS ISTA-DASLab GSQ-RCO IQ3_S both through Strata
21
10
3
177
12,636
Andrew Zhu retweeted
Good news! Qwen4 models will have 1M context 🤩 Except Qwen4-27b, all other models will have 1M token max context. github.com/QwenLM/qwen-code/…
Qwen4-27B !! Today or Tomorrow👀
41
59
12
926
62,462
Andrew Zhu retweeted
New improvements are in! 🫡 github.com/Niko1221/Strata/r… Enjoy 🥳
65
83
32
860
78,501
Back to the Qwen3.8-27B - 18 models tested I ran 18 different quantisations and inferences combinations on my Studio M5 Ultra, you should find this interesting... (more in the thread).
15
8
2
113
146,720
Andrew Zhu retweeted
Aloha!🌺 Meet Ornith-1.5-122B: No Tradeoff Between Strong and Fast Stronger. Faster. Leaner. Ornith-1.5-122B delivers frontier-level capabilities with exceptional inference efficiency. With superior instruction following, stronger task decomposition, and deeper understanding of real-world workflows, it achieves state-of-the-art performance among open-source models of comparable size, while running on a single NVIDIA DGX card!! Open weights under the MIT license. Coding ✅SWE-bench Verified: 0.845 ✅TB2.1: 0.813 ✅CyberGym: 0.800 ✅NL2Repo: 0.512 Agentic ✅HLE: 0.446 ✅BrowseComp: 0.785 ✅MCP-Atlas: 0.746 ✅Toolathlon: 0.708 🤗Hugging Face: huggingface.co/ornith-ai/Orn… We can’t wait to see what you build with Ornith-1.5-122B.
221
497
461
2,747
1,517,721
Andrew Zhu retweeted
Last tests and builds are in progress. A small preview of what's coming in the new version. The release is scheduled for this late evening (European time). I'm open to giving answers to any of your questions below. And sorry for being late!
107
63
37
696
77,341
Andrew Zhu retweeted
You can now run Qwen3.8-Flash-Next on any Mac Mini, Mac Studio or MacBook with Apple Silicon (Even without enough RAM). This is our base performance. From here, it will only get better.
28
31
23
277
89,008
Andrew Zhu retweeted
Next Sushi 🍣 1.2.3 will upgrade Qwen Flash Next to have 20%+ less error with current GDN 16 bits, and with a new 32 bit default, quality will enhance even further. 32bit's tax is ~3% tokens/s, but quality is more important. Avoid that with a flag but still get 20% improvement.
12
2
59
2,380
Andrew Zhu retweeted
Replying to @unbug
Concurrency on strata coming today :)
10
3
1
69
2,422
Andrew Zhu retweeted
Meet Qwen3.8-LiveTranslate, Qwen's next-generation real-time simultaneous interpretation model! 📢 Built on an Interleave architecture, it improves faithfulness, fluency, and conciseness while reducing average lagging (LAAL) from 2.8s to 2.3s across 60 languages. New capabilities: 🙌 - Real-time speaker diarization — distinguishes speakers in multi-party speech and preserves each speaker's voice through more stable voice cloning. - Synchronized bilingual display — source and translation on screen together. - Long-context disambiguation — leverages conversation history to clarify names and terminology for consistent translations. Let's try Qwen3.8-LiveTranslate! 🥳 - Blog: qwen.ai/blog?id=qwen3.8-live… - QwenCloud: qwencloud.com/models/qwen3.8…
112
296
67
2,727
313,732
DeepSeek-V4.1-Flash: it should be called V5, open source, and beat Opus 5.0 by @xhinker generativeai.pub/deepseek-v4…
1
1
84
Andrew Zhu retweeted
great writers on Medium @Beth_Kindig AI scaling is becoming a networking problem as much as a compute problem—@Beth_Kindig breaks down why NVIDIA’s optical push and CPO could reshape the next leg of AI infrastructure. beth-kindig.medium.com/insid… @Beth_Kindig Memory has gone from a cyclical afterthought to a core AI bottleneck—@Beth_Kindig explains why HBM, DRAM, LPDDR and NAND shortages could keep memory strategically important to the AI trade. beth-kindig.medium.com/micro… @pandeyparul Keeping up with AI research is increasingly a workflow problem—@pandeyparul shares a thoughtful approach to combining manual paper reading with LLM assistance without outsourcing the actual thinking. pandeyparul.medium.com/readi… @pandeyparul On-device AI is becoming much more practical—@pandeyparul explores three ways to test models on real phones and why edge deployment is getting easier to experiment with. pandeyparul.medium.com/can-t… @xhinker The AI memory crunch is changing the economics of personal compute—@xhinker lays out a practical 2026 local-AI rig using used server hardware and value-focused GPUs. xhinker.medium.com/build-you… @xhinker Hands-on testing is often more useful than launch-day benchmarks—@xhinker spent days testing MiniMax H3 and breaks down what actually matters across the GPU stack, acceleration and prompting. xhinker.medium.com/minimax-h… @decisionleader Calling layoffs “AI-driven” can hide more than it explains—@decisionleader examines the evidence and incentives behind the narrative and why companies should be careful about blaming workforce reductions on AI. kozyrkov.medium.com/whats-re… @decisionleader Decades of search-engine habits may actually make us worse at prompting—@decisionleader explains why effective AI interaction often requires giving models the context Google trained us to remove. kozyrkov.medium.com/google-s… @Alber_RomGar AI progress may make more sense through the lens of compounding than point forecasts—@Alber_RomGar explains why long horizons and consistent improvement matter more than trying to time every breakthrough. albertoromgar.medium.com/how… @Alber_RomGar More AI agents don’t automatically mean more productivity—@Alber_RomGar looks at the cognitive cost of managing too many parallel AI workflows and why orchestration can become work itself. albertoromgar.medium.com/how… @tomsmith585 The leading AI labs are becoming platform companies rather than startups at the edge of Big Tech—@tomsmith585 looks at how OpenAI, Anthropic and xAI are evolving into much larger competitive forces. medium.com/the-generator/ai-… @tomsmith585 The next phase of AI is moving from software into the physical world—@tomsmith585 looks at robotaxis, robotics and what happens as AI starts acting on the world rather than only generating information. medium.com/the-generator/thi… @jrdothoughts AI scaling is increasingly constrained by physical infrastructure—@jrdothoughts takes power, land and cooling seriously and explores whether distributed AI training could eventually extend beyond Earth. jrodthoughts.medium.com/the-… @YoussefHosni951 RAG becomes much harder once you move beyond a prototype—@YoussefHosni951 walks through the production stack from ingestion and retrieval to reranking, monitoring and evaluation. yousefhosni.medium.com/desig…
3
3
433
这是我目前看到比较优秀的Ai作品了 像爱死机的其中一章…… 一个很完整的短片故事……
141
416
39
2,405
252,370
Andrew Zhu retweeted
やった! DeepSeek-V4.1-FlashがEncoder-Decoderと呼んでいる後半層KV近似の仕組みをQwen3-8Bに適用して、プレフィル時間半分、出力同様というのができた! つまりこの仕組みは、既存のモデルでも、モデル自体をいじらず近似モデルを追加してあげれば、プレフィル時間を半分にすることができる。
24
189
33
1,224
206,186
Andrew Zhu retweeted
欢迎试用 DeepSeek Harness v0.1.5 版本。新版本与 DeepSeek V4.1 Flash 模型训练深度结合,模型在 DeepSeek Harness  的不同配置中都进行了专项训练和优化。 与模型训练深度结合的实验性的 Agent Teams 功能也期待尝鲜试用。 DeepSeek Harness 团队仍在开放招聘中,欢迎加入我们,一起打造 “一切皆插件” 的下一步,一起探索模型与 Harness 共同进化的 Agent 智能上限。
258
99
26
1,993
230,544