Pedro Cuenca

how we implemented Moondream inference on Apple Silicon (spoiler: we don't use MLX) ⬇️ (1/N)

After 5 amazing years at @huggingface I’m happy to share that I have no clothes left without emojis on them

Open source projects like transformers are drowning in AI agent PRs, so we auto-merged everything to see what would happen and share the results. tl;dr: if 100s of agents want to fix something, it’s probably broken. Agent PRs on transformers have quadrupled over the past quarter. We classified and validated 1k PRs (42% features, 39% bugs, 13% docs). The quality distribution is skewed toward noise. But the bug fixes cluster around a small number of hotspots: tokenizer handling, model loading, dtype mismatches, multimodal pipelines. I.e. an underlying problem. When 28 PRs independently flag the same area, that is signal regardless of whether any individual fix is correct. One issue generated 39 near-identical PRs in a day. Each applied the same decorator pattern to a different model file. A maintainer would do the same cognitive work 39 times, so a single combined PR replaces all of that work. We built tooling to cluster, deduplicate, and merge these contributions at scale, then ran an experiment: bulk-merge hundreds of agent PRs into a fork, benchmark it, and see what breaks. Nothing broke. Zero delta across three models on arc_challenge, gsm8k, and hellaswag. The contributors are not adversarial. They lack the context to evaluate whether the agent's output is correct. Check out this blog post, where we dive deep on this pipeline: huggingface.co/spaces/hugging…

This hackathon is going to be wild. Build the fastest Metal kernels for Apple Silicon 🚀 Ben pitched me on this idea in person a couple weeks ago, happy finally seeing it happen!

@ivanfioravanti Taking a look.

@pcuenq @angeloskath It requires trust_remote_code.

MLX Ling-2.6-flash support added! 💪 Here my (preliminary, because I bet @angeloskath will improve performance) context benchmark for the 4bit version running on M3 Ultra (cooking a new version) I created the PR with the amazing transformer_to_mlx skill by @pcuenq and Opus 4.7. Few iterations and it seems 😂 working 100%! Ultra fast model created by @TheInclusionAI. Can't wait to test it with a code harness! Raw results: Ling-2.6-flash-mlx-4bit MLX Benchmark Results Hardware: Apple M3 Ultra, 512.0GB RAM, 32 CPU cores, 80 GPU cores 0.5k pp 632 tg 79 t/s mem 59.4GB kv 0.03GB 1k pp 676 tg 79 t/s mem 59.9GB kv 0.04GB 2k pp 693 tg 79 t/s mem 61.1GB kv 0.04GB 4k pp 704 tg 79 t/s mem 61.1GB kv 0.05GB 8k pp 708 tg 78 t/s mem 61.2GB kv 0.07GB 16k pp 700 tg 77 t/s mem 61.9GB kv 0.11GB 32k pp 678 tg 74 t/s mem 64.5GB kv 0.18GB 64k pp 637 tg 70 t/s mem 69.5GB kv 0.33GB 128k pp 564 tg 63 t/s mem 79.6GB kv 0.64GB Total generated tokens: 1135 Batch TPS: b1 78 b2 123 b4 164 b8 218 b16 307 b32 418 Batch KV : b1 0.04GB b2 0.08GB b4 0.16GB b8 0.32GB b16 0.63GB b32 1.26GB

@ivanfioravanti @angeloskath Also, any feedback on the skill or the process will be very much appreciated 🙏

300,000 AI builders have already added their hardware to HF to instantly see what model they can run locally. To do so, go to huggingface.co/settings/local… and add your hardware specs. You can even show off publicly by adding it to your HF profile! Let's go local AI!

[Hugging Face Machine Learning Club India] It is with immense gratitude and exictement I announce the next speaker for our virtual event, @sarahookr! The date, time, details on how to join will be shared in due time. Share this announcement with you friends, and help make this a great event! 🤗

@ariG23498 It's really beautiful.

@pcuenq I would not have gotten even one task done. Would have just admired the architecture the whole day.

My office today (Escuelas Pías Library, set on the rebuilt ruins of a church and school that burned in 1936 during the civil war)

Humanity's Last Hackathon is NOW OPEN for registration. This is not a normal hackathon. You will be judged on the context, not the code! Use Codex @OpenAIDevs to build and optimize models for local inference (kernels on Max metal). Submit through @GPU_MODE. Climb the leaderboard. Top performers qualify for the final battle. Launches May 4th. Registration is live now.

New open agentic model released by @poolsideai 🔥 Welcome!

In the new release of MLX we are bringing thread safety: def worker(): print(mx.arange(10)) threading.Thread(target=worker) which makes parallel inference easier to implement, a feature that had been driven by projects like omlx/vmlx/vllm-mlx. github.com/ml-explore/mlx…

Mega excited to be officially launching @ElevenLabs Spain today. My home country and one of the fastest growing markets for us. Let's go!!

Shoutout to @0xClandestine for the quick fix! Was literally driving 300KM back from an appointment to fix this. @0xClandestine: “Lemme know if you need anything for your PR”💥 This is why I love this community.

Top 3 trending models of the week on HF: @deepseek_ai @OpenAI & @Alibaba_Qwen!

Today we launch smol-audio A collection of notebooks & scripts to build on cutting-edge local audio models ⚡️ Already in the cookbook: > Fine-tune Whisper / Parakeet / Voxtral / Granite Speech > Fine-tune Audio Flamingo 3 (full + LoRA) > Dialogue TTS with Dia-1.6B > Zero-shot video + audio↔text retrieval with Meta's PE-AV More to come — what would you like to see next? Reply with suggestions.

@ariG23498 @RisingSayak 👀👀

[Hugging Face ML Club India] We are beyond excited for the next virtual event. We host an incredible researcher and more than that an idol of mine (pretty sure of @RisingSayak's as well). They will be talking about the slow death of scaling. I am pretty sure you know who that is, but more information coming soon. Keep your eyes glued to this space. 🤗