Marko Tasic (@mtasic85) - Twitter Profili | Zamantika Mersobahis Locabet

Sabitlenmiş Tweet

We are happy to announce that @TangledGroup has published tangled-nx-sql v0.1.0 which seamlessly adds SQL database as a persistence layer to NetworkX graphs, zero in-memory footprint, NetworkX API compatibility. All examples and test can be found here: pypi.org/project/tangle…

English

1

2

103

Marko Tasic@mtasic85·10m

This is how ideology kills technology. Book example. I really hope more people will move from Debian and Ubuntu now and give chance to others Linux distributions. xubuntu.org/news/releases/…

English

0

3

Marko Tasic@mtasic85·24m

Interesting 🤔 techtrenches.dev/p/the-west-for…

English

0

Marko Tasic@mtasic85·17h

🤏🫘This will probably sound weird, but I migrated from opencode to pi because I needed something lightweight. I got it, but now I need even more lightweight agent/harness. I really like: ``` pi --provider "llama.cpp" --model "Qwen/Qwen3.6-27B" -p "size of README.md" ``` But everything else around it, still makes it too heavy for me. Damn TUI. 🤣

English

0

46

Marko Tasic retweetledi

itewqq@lyq_sqsp·2d

Ubuntu 26.04 is hacked in ~12 hours after it was released. Security in the Age of AI?

DARKNAVY@DarkNavyOrg

Our AI Agent popped a root shell on Ubuntu 26.04 on the first day it was released :)

English

36

338

2.7K

310.8K

Marko Tasic retweetledi

Ubuntu@ubuntu·1d

Ubuntu 26.04 LTS #ResoluteRaccoon has landed. Discover the biggest updates in our announcement: canonical.com/blog/canonical…

English

45

174

1.1K

37.7K

Marko Tasic retweetledi

CG@cgtwts·1d

> be chinese ai labs > while claude and openai are in cold war > kimi dropped k2.6 using deepseek's v3 architecture > the same week deepseek drops v4 using kimi's muon optimizer > 1.6 trillion parameters & 1M context > both match or beat closed models on benchmarks while being 8x cheaper > both build on each other's breakthroughs > keep shipping frontier LLMs with far less or nerfed NVIDA GPUs > and keep them 100% open sourced the real battle is not between models, it's open source vs closed.

DeepSeek@deepseek_ai

🚀 DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length. 🔹 DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models. 🔹 DeepSeek-V4-Flash: 284B total / 13B active params. Your fast, efficient, and economical choice. Try it now at chat.deepseek.com via Expert Mode / Instant Mode. API is updated & available today! 📄 Tech Report: huggingface.co/deepseek-ai/De… 🤗 Open Weights: huggingface.co/collections/de… 1/n

English

49

351

4.5K

251.8K

Marko Tasic retweetledi

Physics In History@PhysInHistory·1d

Understand. Don't memorize. Learn principles, not formulas. - R. Feynman

English

65

1.4K

8K

169.4K

Marko Tasic retweetledi

tetsuo@tetsuoai·1d

how CNNs see images 16 boxes covering the core CNN stack. tensors, filters, feature maps, stride, padding, channels, pooling, receptive fields, mental model

English

13

140

803

30.2K

Marko Tasic retweetledi

Science Simplified@Scivf4·1d

Quantum entanglement isn't instant - scientists have timed It's 'BIRTH' at 232 attoseconds, A quintillion of a second.

English

67

358

2.5K

245K

Marko Tasic retweetledi

China Research Collective@CRC_8341·1d

Yao class (姚班) genius Yao Shunyu (姚顺雨) left OpenAI in late 2025 to return to China, joining Tencent Hunyuan Known in China as one of "The Four Foundation Model Heroes" (基模四杰) Along with Yang Zhilin (Kimi), Lin Junyang (Ex-Qwen), and Tang Jie (Z.ai)

Shunyu Yao@ShunyuYao12

Our goal is to build practical models with comprehensive capabilities beyond open benchmarks. And the only way to do it to co-design with diverse products while scaling solidly. Tencent has the best product ecosystem and a solid, low-ego culture, and we are just getting started!

Filipino

11

133

1.2K

187.5K

Marko Tasic@mtasic85·21h

🟡Current setup, 22.81GiB / 24.00GiB VRAM 🫡35 t/s tg 🟡250 t/s tg, with speculative decoding 🤗Unsloth Qwen3.6-27B Q4_K_M -ctk/v q5_0 🟩Nvidia RTX 3090 24GB / CUDA 12.9.1 CUDA_VISIBLE_DEVICES=%i ./llama-server -hf unsloth/Qwen3.6-27B-GGUF:Q4_K_M -ngl -1 -fa on -fit off --metrics --props --slots --host 0.0.0.0 --port 808%i -dev CUDA0 --temp 0.7 --top-p 0.8 --top-k 20 --min-p 0.0 --presence-penalty 1.5 --repeat-penalty 1.0 --reasoning off --alias "Qwen/Qwen3.6-27B" --api-key sk-pvQ5YvAy -c 262144 -ctk q5_0 -ctv q5_0 --spec-default --no-mmproj-offload -b 1024 -ub 256 CUDA_VISIBLE_DEVICES=0,1,2,3 ./llama-server -hf ggml-org/GLM-OCR-GGUF:Q8_0 -ngl -1 -fa on -fit off --metrics --props --slots --host 0.0.0.0 --port 8090 --reasoning off --alias "zai-org/GLM-OCR" --api-key sk-pvQ5YvAy -c 32768 -ctk q5_1 -ctv q5_1 -np 1 --spec-default --no-mmproj-offload

Marko Tasic@mtasic85

🟠Current setup, 23.37GiB / 24.00GiB VRAM 🤨30 t/s tg 🤔with speculative decoding 250 t/s tg 🤗Qwen3.6-27B Q3_K_M -ctk q8_0 -ctv q8_0 CUDA_VISIBLE_DEVICES=%i ./llama-server -hf unsloth/Qwen3.6-27B-GGUF:Q3_K_M -ngl -1 -fa on -fit off --metrics --props --slots --host 0.0.0.0 --port 808%i -dev CUDA0 --temp 0.7 --top-p 0.8 --top-k 20 --min-p 0.0 --presence-penalty 1.5 --repeat-penalty 1.0 --reasoning off --alias "Qwen/Qwen3.6-27B" -c 262144 -ctk q8_0 -ctv q8_0 --spec-default --no-mmproj-offload CUDA_VISIBLE_DEVICES=0,1,2,3 ./llama-server -hf ggml-org/GLM-OCR-GGUF:Q8_0 -ngl -1 -fa on -fit off --metrics --props --slots --host 0.0.0.0 --port 8090 --reasoning off --alias "zai-org/GLM-OCR" -c 32768 -ctk q5_1 -ctv q5_1 -np 1 --spec-default --no-mmproj-offload

English

0

46

Marko Tasic@mtasic85·23h

🟠Current setup, 23.37GiB / 24.00GiB VRAM 🤨30 t/s tg 🤔with speculative decoding 250 t/s tg 🤗Qwen3.6-27B Q3_K_M -ctk q8_0 -ctv q8_0 CUDA_VISIBLE_DEVICES=%i ./llama-server -hf unsloth/Qwen3.6-27B-GGUF:Q3_K_M -ngl -1 -fa on -fit off --metrics --props --slots --host 0.0.0.0 --port 808%i -dev CUDA0 --temp 0.7 --top-p 0.8 --top-k 20 --min-p 0.0 --presence-penalty 1.5 --repeat-penalty 1.0 --reasoning off --alias "Qwen/Qwen3.6-27B" -c 262144 -ctk q8_0 -ctv q8_0 --spec-default --no-mmproj-offload CUDA_VISIBLE_DEVICES=0,1,2,3 ./llama-server -hf ggml-org/GLM-OCR-GGUF:Q8_0 -ngl -1 -fa on -fit off --metrics --props --slots --host 0.0.0.0 --port 8090 --reasoning off --alias "zai-org/GLM-OCR" -c 32768 -ctk q5_1 -ctv q5_1 -np 1 --spec-default --no-mmproj-offload

Marko Tasic@mtasic85

🟡Current setup, 21.53GiB / 24.00GiB VRAM 🧐40 t/s tg 🧐with speculative decoding 200 t/s tg CUDA_VISIBLE_DEVICES=%i ./llama-server -hf unsloth/Qwen3.6-27B-GGUF:IQ4_NL -ngl -1 -fa on -fit off --metrics --props --slots --host 0.0.0.0 --port 808%i -dev CUDA0 --temp 0.7 --top-p 0.8 --top-k 20 --min-p 0.0 --presence-penalty 1.5 --repeat-penalty 1.0 --reasoning off --alias "Qwen/Qwen3.6-27B" -c 262144 -ctk q4_0 -ctv q4_0 --spec-default --no-mmproj-offload CUDA_VISIBLE_DEVICES=0,1,2,3 ./llama-server -hf ggml-org/GLM-OCR-GGUF:Q8_0 -ngl -1 -fa on -fit off --metrics --props --slots --host 0.0.0.0 --port 8090 --reasoning off --alias "zai-org/GLM-OCR" -c 32768 -ctk q5_1 -ctv q5_1 -np 1 --no-mmproj-offload

English

0

84

Marko Tasic@mtasic85·1d

🟡Current setup, 21.53GiB / 24.00GiB VRAM 🧐40 t/s tg 🧐with speculative decoding 200 t/s tg CUDA_VISIBLE_DEVICES=%i ./llama-server -hf unsloth/Qwen3.6-27B-GGUF:IQ4_NL -ngl -1 -fa on -fit off --metrics --props --slots --host 0.0.0.0 --port 808%i -dev CUDA0 --temp 0.7 --top-p 0.8 --top-k 20 --min-p 0.0 --presence-penalty 1.5 --repeat-penalty 1.0 --reasoning off --alias "Qwen/Qwen3.6-27B" -c 262144 -ctk q4_0 -ctv q4_0 --spec-default --no-mmproj-offload CUDA_VISIBLE_DEVICES=0,1,2,3 ./llama-server -hf ggml-org/GLM-OCR-GGUF:Q8_0 -ngl -1 -fa on -fit off --metrics --props --slots --host 0.0.0.0 --port 8090 --reasoning off --alias "zai-org/GLM-OCR" -c 32768 -ctk q5_1 -ctv q5_1 -np 1 --no-mmproj-offload

Marko Tasic@mtasic85

🟠Current setup, 22.64GiB / 24.00GiB VRAM pattern on 4x @nvidia RTX 3090, @Alibaba_Qwen and @Zai_org models and @UnslothAI and @ggml_org quants on llama.cpp 🧐40 t/s tg 🤯with speculative decoding 400 t/s tg CUDA_VISIBLE_DEVICES=%i ./llama-server -hf unsloth/Qwen3.6-27B-GGUF:IQ4_NL -ngl -1 -fa on -fit off --metrics --props --slots --host 0.0.0.0 --port 808%i -dev CUDA0 --temp 0.7 --top-p 0.8 --top-k 20 --min-p 0.0 --presence-penalty 1.5 --repeat-penalty 1.0 --reasoning off --alias "Qwen/Qwen3.6-27B" -c 262144 -ctk q4_0 -ctv q4_0 --spec-default CUDA_VISIBLE_DEVICES=0,1,2,3 ./llama-server -hf ggml-org/GLM-OCR-GGUF:Q8_0 -ngl -1 -fa on -fit off --metrics --props --slots --host 0.0.0.0 --port 8090 --reasoning off --alias "zai-org/GLM-OCR" -c 32768 -ctk q5_1 -ctv q5_1 -np 1 --no-mmproj-offload

English

0

92

Marko Tasic@mtasic85·1d

🟠Current setup, 22.64GiB / 24.00GiB VRAM pattern on 4x @nvidia RTX 3090, @Alibaba_Qwen and @Zai_org models and @UnslothAI and @ggml_org quants on llama.cpp 🧐40 t/s tg 🤯with speculative decoding 400 t/s tg CUDA_VISIBLE_DEVICES=%i ./llama-server -hf unsloth/Qwen3.6-27B-GGUF:IQ4_NL -ngl -1 -fa on -fit off --metrics --props --slots --host 0.0.0.0 --port 808%i -dev CUDA0 --temp 0.7 --top-p 0.8 --top-k 20 --min-p 0.0 --presence-penalty 1.5 --repeat-penalty 1.0 --reasoning off --alias "Qwen/Qwen3.6-27B" -c 262144 -ctk q4_0 -ctv q4_0 --spec-default CUDA_VISIBLE_DEVICES=0,1,2,3 ./llama-server -hf ggml-org/GLM-OCR-GGUF:Q8_0 -ngl -1 -fa on -fit off --metrics --props --slots --host 0.0.0.0 --port 8090 --reasoning off --alias "zai-org/GLM-OCR" -c 32768 -ctk q5_1 -ctv q5_1 -np 1 --no-mmproj-offload

English

0

88

Marko Tasic@mtasic85·1d

aur.archlinux.org is down 🥶

English

0

19

Marko Tasic retweetledi

ThePrimeagen@ThePrimeagen·1d

I love this I don't say it because I'm somehow anti-ai, I say this because it's destructive for your life. You're going to have disrupted sleep, you're going to have disrupted relationships, your whole life will be lived between thinking of the next prompt and results of the current one. This ain't it king

David Cramer@zeeg

Everyone is slowly coming to this realization, and I assure you, no one is running multitudes of agents overnight. No one that is doing anything of substance at least. There _are_ people pretending to be scientists, or fully caught up in their drug infused AI overdose, that think their slop machines are changing the world. They're not tho, and they're just wasting a bunch of money and compute to create a lot of LoC that will just get thrown away. The state of the art is still "can we even one shot a production quality patch that we wont regret later", and its rarer than you'd expect based on discourse.

English

79

75

2K

163.8K

Marko Tasic retweetledi

Maxime Rivest 🧙‍♂️🦙🐧@MaximeRivest·1d

dspy can feel like a blackbox sometimes. Got UI can completly mitigate that! This is the streaming of a prompt optimization run in dspy. You see every single token as they are generated. The instructions and the evaluations!