Samuel Fajreldines

😨

@ivanfioravanti In my tries, it doens't work well with quantizations. So for many ppl mtp still a better option.

Dflash-MLX 0.1.6 Qwen3.6-27B bf16 with z-lab draft AIME 25 test on M3 Ultra 512GB 83.3% in 1:43 hours! Prompt 235.6 t/s srv Generation 34 t/s srv This has been much faster than my first MTPLX test, I see shorter thinking traces, probably due to bf16 vs 4bit. I'll now try 8bit on both engines and I will raise temp to 0.6.

@faraz0x Discount is not appearing to me :/

ICYMI 👀: SuperGrok Heavy just dropped to $99/month for 6 months. That's 67% off the usual $300. Grok Build beta is live for Heavy subscribers right now. 8 parallel agents, 2 million token context, straight from your terminal. This is the cheapest the most powerful Grok tier has ever been. Worth trying.

@ivanfioravanti If you want to try more MTPLX models: huggingface.co/samuelfaj/coll… :) My favorite: huggingface.co/collections/sa…

MTPLX is keeping up, engine still alive and great results! So far so good! M3 Ultra mtplx-qwen36-27b-optimized-speed AIME 25 6/30 100% 🤞🏻

Fine-tunning my own LLM :D

@DJLougen Awesome work @DJLougen ! This is my new favorite model. I just optimized for MTPLX (disponible in lightning mlx): github.com/samuelfaj/ligh…

Ornstein3.6-27B-MTP-NSC-ACE-SABER is live. Qwen3.6 27B MTP base, put through NSC-ACE, SABER, then Ornstein. MTP preserved -> GGUFs shipping Yes the name is too long huggingface.co/GestaltLabs/Or… huggingface.co/GestaltLabs/Or…

@0xPrajwal_ Give a try and be happy: github.com/samuelfaj/dist…

We pay $200/month for Claude just to hit the limit in under an hour what are we even paying for??💀

@justbyte_ Codex 100% sure

developers, choose one forever

@cjzafir Share more about how you’re doing. I’d love to learn.

I'm building ELMs on top of Open Source Models. I need more compute so I can do more experiments. Where can I go to get more GPUs? Lambda GPUs are mostly unavailable, Runpod is not so consistent either. Any suggestions?

@ivanfioravanti I created github.com/samuelfaj/ligh… to test experimental things that won't fit (for now) in RAPID-MLX

Why in the Apple MLX ecosystem, is everyone building a new engine copying from others, instead of contributing and improving existing ones? 🤔 I the last months we have seen an explosions of new engines, all similar. 😢

@SOSOHAJALAB github.com/samuelfaj/ligh… Is faster :D

This repo is really good for accelerating the mlx github.com/raullenchai/Ra…

MTPLX of Qwen3.6-35B-A3B-NSC-ACE-SABER for lightning-mlx on Apple Silicon. huggingface.co/collections/sa…

@DJLougen This is an excellent work @DJLougen. My main model for now. I've just created MTPLX version: huggingface.co/collections/sa…

Took a bit to do, but here is the Neural Steering Community version of Qwen3.6-A3B-35B model huggingface.co/GestaltLabs/Qw… huggingface.co/GestaltLabs/Qw…

New MTPLX models: Qwopus MLX MTPLX Ornstein MLX MTPLX Qwen MLX MTPLX huggingface.co/samuelfaj/coll… #mlx #llm #mtplx

@Merocle The concept of flexing has being updated.

M5 Max cluster 72 CPU and 128 GPU cores, 512GB unified Ram Each MacBook is connected to all the others with Thunderbolt 5 (120Gbit/s). But I’ll have to use Wi-Fi to connect to the cluster

Uploading new #mtplx models. #mlx #mtp

@antirez What if use JANG instead of 2bit?

DS4 is now called DwarfStar4, since you can put a lot of mass into a tiny space... And in a few minutes it is going to be much better on 128GB Macs because I'l pushing much better 2 bit quants generated with an in-house iMatrix magic recipe.

@AlexJonesax github.com/samuelfaj/ligh… Faster

Qwen3.6-27b absolutely flying on a M5Max with MTP enabled & oMLX inference.

@jun_song And GPT 5.5 is extremely smarter then Opus 4.7

I can't believe people actually still think Anthropic is leading the AI race. Bro they have no compute: • No image gen at all • The model absolutely tanks at peak times • Mythos is still MIA • They had to beg xAI for crazy expensive compute The marketing brainwash is real.