TechPractice

Great news for llama.cpp users: it has added the MTP support, which makes running Qwen 3.6 27B or 35 B much faster. youtu.be/AK9T6qlGErE

@JoeGuglielmucci Thank you!

I couldnt believe it...but, A single 5090 is pulling off 600 tok/s w Gemma4b thanks to DFlash. Local Agents can compete with cloud models, and, win. (Great work: @TechPractice1)

I found that it's even faster when increasing output length. It can go above 700 t/s for 1500 output tok length.

Video link at youtu.be/S_zbHH5Ycs0 How DFlash works: A small draft model proposes N tokens ahead. The main model verifies them all in one forward pass.

Gemma 4 26B running at >600 tok/s on a single RTX 5090. Thanks to DFlash. @zhijianliu_ @googledevs 🧵

Codex is helping me benchmarking vllm + dflash on a RTX5090. Results will be out soon!

The RTX 3090 from 2020 is still competitive with the 5090 for local AI in some tests. I ran Qwen3 27B on 4 GPUs. The results will change how you think about upgrading. 🧵 Full benchmark 👇 youtu.be/QYSmQx55e44?si…

Can Mac run deepseek-v4-flash locally? Find it out at the video: youtu.be/a0GesZXLwfc?si…

Here are simple steps to upgrade CUDA (on Ubuntu): 1. uninstall previous CUDA. sudo apt-get --purge remove "*cublas*" "cuda*" "nsight*" sudo apt autoremove sudo reboot 2. install the new CUDA following Nvidia guide: e.g. for CUDA 12.8: developer.nvidia.com/cuda-12-8-0-do…

New open source dropped from China (1.02T total parameters and 42B active parameters): huggingface.co/XiaomiMiMo/MiM…

How fast is theQwen3.6-27B inference on Mac? I tested 3 sizes: 4bit 8bit and bf16. See them in this video: youtu.be/Y9Ql3P4-peE?si…

@react_dom @johnAGI168 需要第三方平台

@johnAGI168 所以这种宫格图人脸怎么过小云雀检测人脸

两大SOTA模型强强联手 演绎降龙十八掌对阵扫地僧👨‍🦲 GPT-image 2 📷➕Seedance 2.0🎬 好处就是基本上不用写提示词，就输入大概想法 ChatGPT 生成分镜图，然后参考分镜图生成视频即可！ 想要更好效果还是要写好提示词精确指导才行😄

chatgpt gpt-image-2 able to generate multiple images in one pass:

两个最厉害的模型联手 会有什么神奇效果呢？ ChatGPT-image-2 + Seedance 2.0 #seedance

GPT-image-2.0 + Seedance 2.0 = next level video generating. Two steps: 1. ChatGPT txt to image. 2. Image to video model such as Seedance 2.0 used in following video.

video generating in next level. Two steps: 1. ChatGPT txt to image. 2. Image to video model such as Kling 3.0 use in following video. Seedance 2.0 can also work, see next video.

This is NOT real. But Generated by chatgpt's new model.

Best Mac vs best Nvidia GPUs: I benchmarked Qwen3.6-35B-A3B on an RTX 3090, 4090, 5090, and Mac M5 Max using llama.cpp. Same model. Same prompts. Same settings. youtu.be/3vdM73WV7BI

Opus 4.7 released!

RUN Gemma 4 using MLX on Mac - faster than gguf: youtu.be/j_TcABM2qtY