Youssof Al Toukhi
@Youssofal_ • 12,233 subscribers
I’m a 21 y/o hyperfixated dabbler and Forbes featured viral meme building the next biggest mobile game based on memes.
Shorts
Videos
Omg... after using GLM 5.3 flash at 1000 TPS I do not think I can ever go back to anything else.
Youssof Al Toukhi359,816 görüntüleme • 18 gün önce
20 hours later and I’m done! Qwen 3.8 27B @ 73 TPS (peak) on a Macbook pro M5 max. MTPLX V2.7 out now with 3 new models. - Bare Speed: short burst decode, good for chat. - Optimized Speed: higher quality and faster on long coding. - Optimized Quality: 30% slower but super!
Youssof Al Toukhi97,435 görüntüleme • 2 ay önce
I put Qwen 3.6 27B on a B200 for the fun of it. I told Fable to apply my past optimisation principles. Result? 1000 TPS Qwen 3.6 27B NVFP4 (550 TPS average long decode) I used the same principles that got me 250 TPS on 2x 3090s. Now I need to buy a DGX Station 🤪.
Youssof Al Toukhi46,527 görüntüleme • 2 ay önce
Introducing MTPLX V2.12: The fastest way to run local AI now even faster and with more models for lower memory macs and the new M5 Ultra. Upgrades: - 100 TPS Qwen Flash Next sustained on long coding (M5 max). - 2x Prefill speeds Qwen Flash Next: 872 --> 1695 TPS. - Requests with images now use the compiled kernels for max speeds. New Models: - Qwen Flash Next MTPLX Optimized Quality: 8 bit model built for the M5 Ultra and M3 Ultra. - Bonsai 2 27B MTPLX Optimized Speed: Works on 16gb and 24gb macs with compiled kernels and MTP running as fast as possible. 64 TPS. - MiMo V2.6 Qwen 3.5 9B MTPLX Optimized Speed: My personal favourite. Mimos magic distilled into a 9B model easily runnable at high speeds on a 16gb mac. 107 TPS. +40 bug fixes and improvements in real world coding.
Youssof Al Toukhi21,486 görüntüleme • 1 ay önce
Introducing MTPLX V2.11.3: the fastest way to run AI models on your mac now perfectly exact. Qwen 3.8 Flash Next M5 Max Macbook In OpenCode: - Peak Speed: 126.5 TPS. - Average Speed at 100k context: 61 TPS. - Average Speed at 200k context: 50 TPS. I kept hearing on various platforms people complaining about the quality of outputs on MTPLX compared to LM studio. Even though our models themselves were the fastest and highest quality frequently praised something was wrong with the engine itself. After a deep scan running thousands of samples and comparisons I have found and rectified 8 exactness issues. MTPLX the fastest way to run models on your mac is now exact at any temperature. Many other runtimes trade exactness and quality for speed, either utilising the ANE or quantising attention. MTPLX will always peruse speed as long as it does not diminish exactness and quality. Here you can see it answering what the Navier stokes equation is and make a flappy bird game in a single chat at 80+ TPS without breaking a sweat!
Youssof Al Toukhi20,002 görüntüleme • 1 ay önce
God I am too addicted to Astra Ultrafast. This type of prompt would normally take 30 minutes. With ultrafast it took 5. But it also drew 7% weekly usage in 5 minutes. Each plan can last me maybe half a day if I am smart. I am genuinely considering getting 14 $500 plans to last me the week. Yes, I am aware it is irrational. Maybe when factoring in resets I could last with 10 plans or so. Ugh, I would prefer Claud fast mode, but it is only available via API and its half the TPS and 3x the verbosity. Though the outputs are better.
Youssof Al Toukhi13,364 görüntüleme • 26 gün önce
MTPLX V2.9 is OUT! - 30% faster TPS in CLI & 60% in app. - All MTPLX models retuned to be smaller AND faster at the same quality (re-download!) - 80% lower CPU utilisation in app. - 95% less streaming freezes. Here is our bare speed model using session bank to reach 175 TPS:
Youssof Al Toukhi25,133 görüntüleme • 2 ay önce
MTPLX V2.10 is out! I put a lot of work into this one. Highlights: - Qwen 3.8 Next Support Added: Bare Speed (4 bit) and Optimised Speed (4 bit/8 bit) models added running at 50 - 80 TPS sustained at high contexts. (expect speed improvements in the coming days) - Qwen 3.8 Next mmap: The 32 GB n-gram table streams from SSD via mmap instead of RAM. Both models fit a 96 GB Mac with 0 speed impact. - Qwen 3.8 27B Decode & Prefill boost EVERYWHERE: +15% TPS at 3k context, +29% at 88k, +54% at 147k. Prefill +41%. - 48 GB Macs went from 3 to 4 TPS in swap @ 30k context to 33 TPS. - KV Cache has been solved!: 8 bit and 4 bit cache now barely impacts speed instead of a 50% decline. - Coding agents wall time optimisations: file task dropped from 150s to 44s, and mid-session first token went from ~2s to 0.11s. - CLI UX improvements + 34 bug fixes. My thoughts on Qwen Next? Incredible model. Enormously better at design, and creative tasks with scary good vision. It created what I think is the best flappy bird. maybe even better than the Fable version. It had various difficulties and skins, moving pillars, and upgrades. It also had the most memeified look. Incredibly impressed. However, it is slightly worse than 27B at having code just work straight away. But it's personality is also far better than 27B. Here is the Flappy bird it made below! Made with Qwen Next Optimized Speed one shot@ 57 TPS:
Youssof Al Toukhi21,228 görüntüleme • 2 ay önce
Introducing MTPLX V1: The fastest and simplest way to run MTP compatible models on your Mac. - New Swift based app. 2x speed increase without bloat - Easy OpenCode, Hermes & Pi integration - Convert your own MTP models with forge And more try now at:
Youssof Al Toukhi31,741 görüntüleme • 4 ay önce
Qwen 3.8 27B vs Fable, Kimi K3, GPT 5.6 Sol Pro on Flappy Bird. Prompt: “Make the ultimate adorable cute and beautiful flappy bird game in HTML.” Qwen 3.8 27B Xhigh INT8: - Video 1 (Puff) - 8m 11s - 160 TPS (2x3090) Fable 5 Max: - Video 2 (Flappy Fluff) - 20m 21s Kimi K3 Max Fireworks Fast: - Video 3 (Fluffy Flap!) - 4m 40s - 111.3 TPS GPT 5.6 Sol Pro: - Video 4 (Flappy Boom) - 19m 7s Overview: Personally, I think Fable is the best. It feels extremely polished, sharp and stands out with small things like the dizzy effect on death. However, 27B is freakishly close. Never would I have guessed 2 months ago this would have come from a local model. 27B actually beats Fable on certain aspects like the 3D tilting ui buttons on death. Kimi K3 felt the best to play mechanics wise. It felt smooth with the perfect balance of difficulty and ease to play. Also, it was the only one that had music. 5.6 Sol Pro was hilariously bad in all aspects. I tried 5.6 xhigh multiple times and the outputs were even worse.
Youssof Al Toukhi12,761 görüntüleme • 2 ay önce
Managed to get QWEN 3.6 27B running in Cursor with localhost (no Ngrok) at 120 - 140 TPS on 2x 3090s.
Youssof Al Toukhi21,221 görüntüleme • 5 ay önce
Daha fazla içerik yok.