le fromage

have you tried the fan on/off that @Youssofal_ recommended to use with mtplx ? : mtplx max --on mtplx max --off it sets/resets the fan to max thru thermalForge github.com/ProducerGuy/Th… I've tried it on my M4 Max, it works. for maximum speed of LLMs works better for some models than others

Interesting video of M5 Max, on impact of Low, Automatic and High power modes on inference. - No external monitor attached - Model not relevant, but it's DS4 Flash Q2. Results: - Low ~25W ~12 toks/s - High ~120W ~ 32 toks/s - Automatic varies from 40W ~14 to 90W ~29 in relation to the fan speed and temperature of the Mac. If you really want to push your MacBook to the max, High Power mode and no external monitors, with them I see a very strange behavior that I'm investigating 🧐

@ivanfioravanti please try the new dflash-mlx v0.1.5 ... beats everything github.com/bstnxbt/dflash…

MLX Qwen3.6-27B benchmark, here you can clearly see the difference in TFLOPs and Memory Bandwidth 🥇 M3 Ultra faster decode due to memory bandwidth 🥈 M5 Max faster prefill due to Neural Accelerators Wondering why memory usage is slightly higher on M3 Ultra 🤔

@JungKook941892 grokipedia.com/page/Et_si_tu_…

This is the french rendition youtube.com/watch?v=0fYr5v… from grokipedia : "Et si tu n'existais pas" is a French romantic ballad recorded and popularized by American-born singer Joe Dassin in 1975. It appears as the opening track on his ninth studio album, Joe Dassin (also known as Le Costume blanc), released that year by CBS Records.[1] The song's lyrics, which express profound devotion by imagining a purposeless existence without the beloved, were adapted into French by Pierre Delanoë and Claude Lemesle from the original Italian composition "Oasis," written by Toto Cutugno, Pasquale Losito, and Vito Pallavicini. The music features a gentle orchestral arrangement that underscores its sentimental tone, making it one of Dassin's enduring signature pieces.Oasis est finalement enregistrée par Albatros en 1998 et a été incluse dans l'album Il meglio la même année[5],[6].

❤🌹🍀 One of the most beautiful French songs ever written

@bstnxbt @jundotkim awesome speed!!!

dflash-mlx v0.1.5 is out. Big runtime + serving release since v0.1.4: ► one CLI: serve / generate / benchmark / doctor / profiles / models ► public runtime suites: smoke / humaneval / gsm8k / math500 / longctx ► L1 RAM + L2 SSD snapshot cache, inspired by oMLX (@jundotkim) ► LCP prefix matching, byte budgets, max snapshot policy ► experimental --target-fa-window for long-context Qwen3.5 / 3.6 ► diagnostics basic/full: cache events, request events, cycle timings, memory waterfall ► cleaner server loop + model-provider flow + target- backend seam Shoutout @0xClandestine for draft quantization, Metal MMA draft kernels, and recurrent masked-state fixes shipped in this build. Main point: dflash-mlx now has a cleaner base for runtime work and external PRs. The next target is a FlashAttention-style prefill kernel for long-context runs on Apple Silicon. github.com/bstnxbt/dflash…

Top Row (left to right): 1. Rustichella d'Abruzzo (brown package with egg pasta) 2. Garofalo (blue package with pasta in a net-like design) 3. Rustichella d'Abruzzo (red/black cube box, premium Italian pasta) Middle Row (left to right): 1. Rummo (white package, labeled "RUMMO") 2. La Molisana (yellow/orange package) 3. La Molisana (white package) Bottom Row: 1. De Cecco (blue package) These are all well-known Italian pasta brands. Rustichella d'Abruzzo and Garofalo appear in premium/artisanal packaging, while Rummo, La Molisana, and De Cecco are popular everyday Italian brands.

Complètement d’accord avec ce classement, je valide ! De toute façon, autant choisir des bonnes, c’est la seule chose qu’on pourra se payer bientôt, avec les œufs.

@ivanfioravanti getting tg 727: 60 tokens/sec! 🔥 on M4 Max Nice!

MLX: new Arcee AI Trinity Mini model running on M3 Ultra 512 in 4bit! 112 tokens/sec! 🔥 mlx_lm.generate --model ~/Trinity-Mini-4bit --prompt "Write a poem on LLMs" -m 2048 PR ready, models on mlx-community 🚀

Retweet if you want llama.cpp added here 👇

@_akhaliq @grok summarize paper in 3 paragraphs

Nvidia just dropped FFN Fusion Rethinking Sequential Computation in Large Language Models

Is this the future of AI browser agents? 👀 WebGPU-accelerated reasoning LLMs are now supported in Transformers.js! 🤯 Here's MiniThinky-v2 (1B) running 100% locally in the browser at ~60 tps (no API calls)! I can't wait to see what you build with it! Demo + source code in 🧵👇

What she said.

The new, updated, version just dropped. This book should be on everyone’s reading list for 2023.

The ATA is pleased to report that Shaowen Bao, PhD, an ATA Scientific Advisory Committee member, University of Arizona in Tucson, is investigating commonalities in people who reported sudden or increased tinnitus after a COVID-19 vaccination. abc15.com/news/local-new…