#mlx — Public Fediverse posts

Локальный агент для диагностики инфраструктуры

В статье описаны результаты, которые получил в поисках ответа на вопрос "можно ли решать реальные задачи диагностики и исправления проблем инфраструктуры на слабом MacBook в агентском режиме (да, но)".

https://habr.com/ru/articles/1033614/

#mlx #llmагент #local_llm #local_ai #agent

One of the coolest things of making an ultra fast, local, TTS system, is that you can give real time voices to your AI systems. Play with a demo here. It's a bit like the Realtime API from OpenAI, but free instead of 10 cents per minute. Play with it on the website to chat with Speaklone in real time. Why pay for voice?
https://speaklone.com
#iOS #macoS #indiedev #mlx #apple

- Cohere Transcribe automatic speech recognition model supports 14 languages with impressive benchmarks: https://cohere.com/blog/transcribe https://huggingface.co/CohereLabs/cohere-transcribe-03-2026 MLX port already: https://github.com/Blaizzy/mlx-audio

- Distributed ML training across MacBooks via MLX + Airdrop, cool! https://github.com/swarnim-j/grove

- Rumours: iOS 27 will open Siri to run any AI service (Bloomberg) + Anthropic acknowledges testing 'step change' level model after 'leak' (fortune.com)

#AI #AINews #ASR #cohere #mlx #rumours #siri #anthropic

#Apple should’ve ignored the (pseudo) AI hype

Continue #NeuralAccelerator hardware & #MLX software development, enable running useful LLM locally

Partner with Steam, make running #Games on #macOS & porting to #iOS trivially easy

Embrace a “local first, intermittent connections, eventually consistent” view of the future

Be an alternative to the “cloud first, always on, always connected” future everyone else in trying to sell

Focus on #HomeAutomation

#UnpopularOpinion #MachineLearning

Hi everyone. I am excited to announce that we have released an exciting command line tool called PerspectiveCLI. This tool allows anyone using the Mac terminal to chat with Apple Foundation Models or MLX Community models. You can download it from our Github page, and I encourage anyone to contribute to the project. https://github.com/Techopolis/PerspectiveCLI, #iOSDev #AppleFoundationModels, #MLX, #PerspectiveIntelligence

LM Studio 0.4.0
https://lmstudio.ai/blog/0.4.0
#ycombinator #local_ai #local_llm #gpt_oss #on_device_ai #run_local_ai #LM_Studio #Llama #Gemma #Qwen #DeepSeek #llama_cpp #mlx

Thunderbolt‑5‑Cluster: RDMA macht KI-Berechnungen auf dem Mac deutlich schneller
Mit macOS Tahoe 26.2 bringt Apple RDMA-Unterstützung über Thunderbolt 5 auf den Mac und öffnet damit neue Wege für KI‑Berechnungen im Cluster. Ein Praxistest mit vier Mac Stu
https://www.apfeltalk.de/magazin/feature/thunderbolt%e2%80%915%e2%80%91cluster-rdma-macht-ki-berechnungen-auf-dem-mac-deutlich-schneller/
#Feature #KI #KIForschung #LargeLanguageModels #MacStudio #MacOSTahoe #MLX #RDMA #Thunderbolt5

Mit dem aktuellen Update der #LMStudio #MLX Runtime (0.36.1) laufen seit heute auch die Ministral- und Devstral-Modelle im entsprechenden Format.

ministral-3-14b-reasoning liefert dabei auf meinem 2022er MacBook Pro M1 brauchbare 16tok/sec - dem #LLM beim "Denken" zuzugucken ist dabei recht amüsant: Im Vergleich zu anderen Reasoning-Modellen finde ich es irgendwie sympathisch "verkopft" und unentschlossen.

#ministral #devstral

Apple – So viel schneller laufen lokale KI-Modelle mit dem M5-Chip
Apple demonstriert den Leistungszuwachs des neuen M5-Chips bei der Ausführung von lokalen KI-Modellen auf der eigenen MLX-Plattform. Der Vergleich zum Vorgänger M4 bietet Einblicke in die nächste Generation der Apple-Prozessoren.

Leistungssprung b
https://www.apfeltalk.de/magazin/news/apple-so-viel-schneller-laufen-lokale-ki-modelle-mit-dem-m5-chip/
#KI #News #Apple #KI #LokaleSprachmodelle #M5 #Mac #MLX #Prozessor

Một script Python đơn giản hỗ trợ chép âm thanh micro trực tiếp bằng mô hình parakeet-tdt-0.6b-v2/3 trên MLX, tự động sao chép và dán. Nhấn tổ hợp phím để bật/tắt. Tác giả: @fullbridgerecctifier. Cảm ơn nguồn chia sẻ!

#Python #Transcribe #MLX #parakeet #SpeechToText #ScriptĐơnGiản #TríTuệNhânTạo

https://www.reddit.com/r/LocalLLaMA/comments/1oiofso/single_python_script_for_parakeettdt06bv23_live/

💡 Apple FastVLM funziona offline nel browser: zero latenza e privacy totale nei video IA

https://gomoot.com/apple-fastvlm-funziona-offline-nel-browser-zero-latenza-e-privacy-totale-nei-video-ia/

#ai #apple #appleglasses #ar #blog #fastvlm #llm #mlx #news #picks #smartglasses #tech #tecnologia #webgpu

Apple's MLX adding CUDA support

https://github.com/ml-explore/mlx/pull/1983

#HackerNews #MLX #adding #CUDA #support #machinelearning #CUDA #AppleTech #developernews

Cobbled together an #ExoLabs cluster to fuck around with #devstral a bit, since it's kinda too big for my M3 Max daily driver. While in the process of bringing up nodes the model hit a bug in the #MLX #Python module that deals with inference model sharding related to passing around MLX vs Numpy data structures.

For shits and giggles and also not being a top-tier #Numpy data structure debugging guy I asked Devstral to look at the bug and figure out a fix. After one wrong turn it came up with a fix which I applied to the other nodes and now it's happily sharding the bigger Devstral models. Not sure about vibe coding as a social contagion but from a “How close are we to #Skynet”-perspective I think we're cooked, chat.

Anyway enjoy your Memorial Day weekend 🎉

Figure 1. A very heterogeneous Exo cluster.

How to vibe code for free: Running Qwen3 on your Mac, using MLX
https://localforge.dev/blog/running-qwen3-macbook-mlx
#ycombinator #Qwen3 #MLX #macOS #Apple_Silicon #Local_LLM #Localforge #Free_Code_Generation #Ollama #Local_AI #LLM_Agent

Does anyone know a way to run a very large #GGUF or #MLX pre-trained #AI model using sharding if it won't fit into unified memory? Speed isn't the goal; just loading. I tried a 250GB model with 72GB VRAM + 24GB RAM using the llama.cpp Metal-enabled runtime, but it didn't work in #LM_Studio even with "keep model in memory" off and "try mmap()" on.

Seems like swap or partial loading should be possible, esp. using #macOS dynamically-sized compressed swap. Thoughts?