transcribe.cpp

Evidence (20)

• 4Hugging Face Blog·9/22/2026open_source
  Jun Kim joins Hugging Face to support the MLX community

  Hugging Face announced that Jun Kim, creator and maintainer of oMLX, has joined the company to help support the MLX community. The post says this will provide more stability and long-term funding for oMLX while keeping it Apache 2.0 and maintaining Jun’s leadership. Hugging Face also said it wants to help streamline bringing new Transformers model definitions to MLX implementations.

• 3Hacker News·9/21/2026research
  Transformer Visualizer Explainer

  This interactive article explains how transformer models work through visual illustrations of key components such as attention and token processing. It is an educational resource aimed at making the architecture easier to understand rather than announcing a new model or product.

• 7Hugging Face Blog·9/21/2026research
  Pruning LLMs by Mapping Block Removal to Ising Optimization

  Hugging Face highlights a research approach that treats transformer block removal in large language models as a constrained binary optimization problem. By mapping block selection to an Ising glass, the method uses pairwise block interactions to predict pruning quality and reports strong results on Llama-3.3-70B-Instruct at 50% compression.

• 5Hacker News·9/12/2026research
  Reverse-Engineering Apple's M1 Neural Engine

  This article revisits the reverse engineering of Apple's M1 Neural Engine and maps its internal architecture, including compute, datapath, scheduler, memory, and execution model. It argues that the ANE was optimized for CNN-era workloads and that Apple’s newer M5 design signals a shift toward transformer-oriented GPU integration rather than a standalone NPU.

• 7Hacker News·9/9/2026research
  GPT-6 Astra, Looped Transformers, and Hidden Reasoning

  The article reviews early impressions of OpenAI’s GPT-6 Astra, describing it as a strong model with especially notable performance on math, coding, and 3D/rendering tasks. It then explains looped or recurrent transformer blocks and discusses whether they relate to hidden chains of thought, alongside recent research on the topic.

• 3Hacker News·9/8/2026open_source
  Show HN: LLM Attention Visualizer

  A Hacker News post introduces an interactive tool for visualizing attention patterns in large language models. The project appears aimed at helping users inspect how tokens relate to each other during model inference and better understand LLM behavior.

• 5Hacker News·9/4/2026research
  Why LLMs Are More Than Next-Token Predictors

  The article argues that describing LLMs only as next-token predictors is an incomplete mental model. It explains that while pre-training fits this description, modern post-training methods such as RLVR also teach models from sequences they generate themselves and the rewards those sequences receive. The post uses chess engines as an analogy to show the difference between imitation of existing data and learning through exploration.

• 8Hacker News·9/1/2026research
  Small Transformer Trained in 1.5 Hours Beats Many LLMs on ARC

  The author reports training a small transformer from scratch in about 1.5 hours on a 5090 GPU and claims it matches or beats several larger models on ARC-style evaluation. The updated approach adds architectural and training changes such as SwiGLU, RMSNorm, fewer augmentations, and using ARC-2 non-overlapping tasks while filtering duplicates to avoid data leakage.

• 5InfoQ·8/31/2026research
  Running AI Workloads in the Browser at the Edge

  James Hall outlines why AI workloads are moving from cloud services to local edge devices, with a focus on browser-based execution. He describes using WebGPU, Transformers.js, and DuckDB to improve performance, reduce privacy risk, and support more rigorous evaluation of browser inference.

• 6Hacker News·8/20/2026research
  On-device piano autocomplete model runs in real time on iPhone

  A developer trained a 125M-parameter transformer to autocomplete piano performances in real time on-device, reportedly reaching about 108 notes per second on an iPhone 15. The work highlights the importance of MIDI representation, aggressive data cleaning, and DPO post-training for improving musical continuation quality.

• 8The New Stack·8/5/2026research
  Google reshuffles DeepMind leadership as four senior engineers launch Discovery Loop

  Google is reorganizing DeepMind leadership, with Demis Hassabis moving to a chair/chief scientist role and Koray Kavukcuoglu taking broader control of Gemini model development and related products. At the same time, Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le are leaving to found Discovery Loop, a public-benefit company focused on using AI to automate scientific and engineering research, with Google remaining an investor and cloud provider.

• 7Hacker News·7/19/2026open_source
  transcribe.cpp launches as a ggml-based speech-to-text library

  transcribe.cpp is a new v0.1.0 transcription library built on ggml, aiming to support a wide range of modern speech-to-text models with GPU acceleration. The author says each model is numerically validated and WER-tested against reference implementations, and the project includes bindings for Python, JavaScript/TypeScript, Rust, and Swift/ObjC.

• 5Hacker News·7/14/2026research
  LeMario Trains a JEPA World Model on Super Mario Bros

  A developer reproduced the LeWorldModel architecture and trained a Joint-Embedding Predictive Architecture (JEPA) model on Super Mario Bros. The write-up reports that the model could predict short-horizon futures and support limited reward-free planning, but failed to reliably make long-term progress through the game.

• 8Hugging Face Blog·7/8/2026framework_update
  Transformers backend in vLLM reaches native speed

  Hugging Face says the transformers modeling backend in vLLM now matches or exceeds the throughput of hand-written native vLLM implementations for several Qwen3 models. The update lets model authors run Hugging Face Transformers models in vLLM with a single flag while keeping standard parallelism and serving workflows intact.

• 7Hacker News·7/2/2026research
  Single Transformer Layer Matches Full RL Training

  A research paper on arXiv reports that a single Transformer layer can match the performance of a full-parameter reinforcement learning training setup on the studied task. The result suggests that much of the benefit may come from a smaller portion of the model than expected, though it is still an early research finding.

• 6Hugging Face Blog·6/29/2026research
  DiScoFormer unifies density and score estimation in one transformer

  Hugging Face highlights DiScoFormer, a transformer-based model that estimates both a distribution’s density and score from a sample in a single forward pass. The approach uses a shared backbone with separate heads and a consistency loss, and the authors show that attention can reproduce kernel density estimation as a special case while extending it to higher-dimensional settings.

• 7Hugging Face Blog·6/24/2026framework_update
  NVIDIA NeMo AutoModel boosts MoE fine-tuning in Transformers v5

  Hugging Face describes how NVIDIA NeMo AutoModel builds on Transformers v5 to improve fine-tuning performance for Mixture-of-Experts models. Using the same from_pretrained() API, it adds Expert Parallelism, DeepEP all-to-all dispatch, and TransformerEngine kernels, delivering higher throughput and lower GPU memory use.

• 5Hugging Face Blog·6/23/2026framework_update
  Transformers.js tests proposed Cross-Origin Storage API for shared caching

  Hugging Face describes experiments in Transformers.js with the proposed Cross-Origin Storage API to reduce duplicated downloads and storage across different origins. The post shows that model files and shared WebAssembly runtime assets can be cached once and reused, instead of being fetched again for each site using the same resources.

• 5Hugging Face Blog·6/18/2026research
  Hugging Face asks whether alternatives can beat LoRA

  Hugging Face published a post examining parameter-efficient fine-tuning methods and asking whether LoRA is truly the best option. The article explains why PEFT matters, highlights LoRA’s dominance in usage, and suggests that its popularity may be reinforced by visibility and ecosystem support rather than performance alone.

• 7Hugging Face Blog·5/19/2026model_release
  OlmoEarth v1.1 cuts satellite model compute costs up to 3x

  Allen Institute for AI released OlmoEarth v1.1, a new family of Earth observation models built to improve efficiency while preserving performance on benchmarks and partner tasks. The update focuses on shorter token sequences for satellite imagery processing, enabling lower compute cost, faster inference, and cheaper deployment at scale.