Shippy is a maritime AI agent built for high-stakes decisions, where the wrong answer has real impacts. The article covers its architecture—soul, skills, config—and key design decisions like using a deterministic CLI for API access, sandboxed hosting for user isolation, and a custom evaluation system that scores the whole agent against live data. Lessons learned and future plans are also discussed.
Shippy’s architecture consists of a soul (system prompt), skills (Markdown files), and config, enabling versioned and auditable deployments.
A dedicated CLI abstracts complex API calls, reducing errors and ensuring predictable tool use.
Model routing in AI agents is more complex than it seems. It is not a classification problem but a systems optimization problem involving cost, complexity, and latency. The article shares three key challenges and explains IBM Research's optimization-based approach.
Actual cost depends on caching behavior, not just model pricing.
Task complexity is often invisible at routing time, and routers must balance multiple objectives.
Existing benchmarks suggest voice AI is nearing human-level performance but real-world conversations tell a different story. Hume AI introduces Real World VoiceEQ, a benchmark evaluating over 40 voice models across 15+ dimensions and 60+ metrics, based on over 1 million human ratings. Key findings include: progress is becoming specialized, models are better at speaking than listening, traditional benchmarks overestimate real-world performance, and human evaluation remains essential.
Real World VoiceEQ evaluates 40+ voice models on 15+ dimensions and 60+ metrics using over 1 million human ratings.
Voice models show gaps between speaking and listening abilities; many remain transcript-driven, missing paralinguistic cues like tone and emotion.
NVIDIA emphasizes the importance of open data and synthetic data for building agentic AI, highlighting data inspectability, quality, and trust. The article details Nemotron datasets, the Prompt Atlas visualization tool, and the use of synthetic personas for local diversity.
Synthetic data is crucial for scaling agentic AI while protecting proprietary signals.
NVIDIA's Nemotron open datasets span over 10 trillion pretraining tokens and millions of post-training samples.
The transformers vLLM backend is now as fast (or faster) than custom vLLM implementations for many LLM architectures. Model authors can automatically leverage their transformers implementations to get ultra fast vLLM inference, for free.
Transformers vLLM backend matches or exceeds native vLLM throughput on Qwen3 4B, 32B, and 235B MoE models.
Dynamically applies inference-specific layer fusions at runtime using torch.fx and ast to match custom code speed.
Hugging Face and Amazon SageMaker AI announce a deep-link integration enabling one-click transition from model discovery to SageMaker Studio. The integration pre-configures permissions, surfaces GPU quotas, and supports model customization and deployment, streamlining the path from inspiration to enterprise deployment.
One-click deep link from Hugging Face model page to SageMaker Studio with pre-loaded model and configured environment.
New Studio environments automatically include full permissions for fine-tuning, training, notebooks, and endpoint deployment.
SkyPilot and Hugging Face collaborate to allow users to store models and datasets on the Hub while running compute on any cloud without egress fees for reads.
Mount Hugging Face Buckets or repos into SkyPilot jobs via hf:// URLs and HF_TOKEN
Supports 20+ clouds, Kubernetes, and on-prem clusters
LeRobot v0.6.0 introduces world model policies (VLA-JEPA, FastWAM, LingBot-VA), new VLAs (GR00T N1.7, MolmoAct2, etc.), reward model API (Robometer, TOPReward), six new simulation benchmarks, and a deployment CLI with DAgger corrections, depth sensing, automatic language annotation, up to 2x faster data loading, cloud training, and a leaner install—all aimed at closing the robot learning loop.
Three new world model policies enable robots to imagine future states before acting.
New VLAs include GR00T N1.7, MolmoAct2, EO-1, Multitask DiT, and EVO1, with fine-tuning and deployment support.
This article details the data pipeline behind PRX, a 7B text-to-image model. Key aspects include assembling a diverse pre-training dataset from public and internal sources, using long accurate captions generated by a VLM, and employing Lance for dataset building and MDS for streaming. The team explains their choice of JPEG encoding at quality 92, on-the-fly text latent computation, and lessons learned about data fragmentation.
Pre-training data is assembled from a mix of public and internal datasets, re-captioned with a VLM for consistency.
Long, faithful captions are crucial; they turn imperfections into controllable attributes.
Hugging Face's Kernels project, aimed at standardizing custom kernel packaging, distribution, and consumption, has undergone a major redesign. This post summarizes key updates: a new 'kernel' repository type for better discoverability; enhanced security through trusted publishers and code signing; revamped CLIs with clearer separation of concerns; expanded framework support including Torch Stable ABI and Apache TVM FFI; a foundation for agentic kernel development; and miscellaneous improvements like simplified environment setup and compatibility checking.
New 'kernel' repository type on the Hub allows users to filter by accelerator, OS, and backend version. All kernels are now listed at https://huggingface.co/kernels.
Security improvements: by default, only trusted publishers' kernels are loaded; optional code signing with Sigstore's cosign using ephemeral keys protects against credential compromise.
Hugging Face and Cerebras have collaborated to create a real-time voice AI system powered by Gemma 4, achieving dramatically lower latency through an open modular architecture. The pipeline integrates Nvidia's speech recognition, Cerebras's fast inference, and Alibaba's text-to-speech, and is already deployed in over 9,000 Reachy Mini robots.
Hugging Face and Cerebras debut a real-time voice AI demo using Gemma 4 with ultra-low latency.
The system uses an open cascaded architecture: speech input → ASR → model inference → TTS → spoken response.
IBM Research introduces ScarfBench, an open benchmark for evaluating AI agents on cross-framework migration tasks in Enterprise Java. The benchmark includes 34 applications, 102 framework implementations, and 204 migration tasks. Current top agents achieve less than 10% behavioral success, highlighting the difficulty of preserving behavior during migration.
ScarfBench evaluates AI agents on framework migration between Spring, Jakarta EE, and Quarkus, requiring build, deployment, and behavioral validation.
The benchmark comprises 34 applications, ~2,000 source and test files, and 1,331 expert-written tests.
This article argues that specialization is an inevitable consequence of finite resources and selection pressure, drawing from optimization theory (No Free Lunch theorems), evolutionary biology, competitive markets, and machine learning. It distinguishes specialization from domain knowledge and addresses the Bitter Lesson, concluding that scaling does not eliminate the need for focused systems.
The No Free Lunch theorem implies that no algorithm is universally optimal; specialization trades breadth for fit.
Biology and markets show that limited resources drive niche specialization and concentrated strategy.
Every Eval Ever (EEE) and Hugging Face Community Evals are now intercompatible, allowing cross-posting and interpretation of evaluation results with links to open models, leaderboards, and a unified standardized metadata store.
EEE and Hugging Face Community Evals now interoperable, enabling cross-posting of evaluation results.
EEE provides a unified JSON schema for recording evaluation details, including runner, model, settings, etc.
DiScoFormer is a transformer that estimates both density and score of a distribution from a set of data points in a single forward pass without retraining. It uses cross-attention, a shared backbone with two heads, and a consistency loss to adapt to new distributions. It significantly outperforms KDE in high dimensions.
Estimates density and score simultaneously without retraining.
Leverages consistency loss for self-adaptation to out-of-distribution data.
Spin up a private, OpenAI-compatible LLM endpoint on Hugging Face infrastructure with a single command — no servers to provision, no Kubernetes, pay-per-second. Covers the full process from launch, querying, cleanup, scaling to larger models, creating a chat UI, SSH debugging, and using as a coding agent backend, with a comparison to Inference Endpoints.
Use the 'hf jobs run' command with the vLLM Docker image and --expose 8000 to run a vLLM server on HF Jobs.
Endpoints are authenticated via Hugging Face tokens, requiring read access to the job's namespace, and support querying via curl or OpenAI Python client.
Ai2 compares its 7B transformer Olmo 3 and hybrid Olmo Hybrid, finding the hybrid excels on content words (nouns, verbs, adjectives) and tokens requiring context, but loses advantage on repeated tokens and closing brackets. Token-level loss filtering reveals architectural differences.
Hybrid models predict meaningful tokens (e.g., content words) better, but not repeated tokens.
Hybrids replace some attention layers with recurrent layers, which have fixed-size memory suited for tracking sequential state.
NVIDIA NeMo AutoModel builds on HuggingFace Transformers v5, adding Expert Parallelism, DeepEP fused all-to-all dispatch, and TransformerEngine kernels to achieve 3.4-3.7x higher training throughput and 29-32% less GPU memory for fine-tuning MoE models, with no API changes.
NeMo AutoModel subclasses AutoModelForCausalLM, requiring only one import line change for performance gains.
On a 550B model, Expert Parallelism enables full fine-tuning across 16 nodes of H100s, where Transformers v5 runs out of memory.
CUGA is IBM's open-source agent harness that handles the plumbing of building agentic apps, leaving developers to write only a tool list and a prompt. This article walks through one example — an IBM Cloud advisor app — and explains how CUGA's planning, reflection, and policy system enable robust, production-ready agents.
CUGA abstracts away orchestration, state, and tool calls, letting developers focus on tools and prompts.
The cuga-apps repository contains two dozen single-file apps, each a working example that can be read and copied.
This article explores the Cross-Origin Storage (COS) API proposal, which enables web apps to share large files (like AI models and Wasm runtimes) across origins using cryptographic hashes instead of URLs. Using Transformers.js as an example, it highlights the redundancy caused by current cache partitioning and how COS addresses it with hash-based identification, flexible access control, and integrity verification.
Current browser caches are partitioned by origin, leading to redundant downloads of shared AI resources across different apps.
The Cross-Origin Storage (COS) API identifies files by cryptographic hash, enabling cross-origin sharing.
Hugging Face revamped the release process for huggingface_hub, using AI and open tools to ship weekly releases instead of monthly, while keeping a human in the loop for final review. The new pipeline costs about $0.25 per release and has improved release note quality and discovery of integration issues.
Release cadence improved from 4-6 weeks to weekly
AI drafts release notes, but deterministic verification ensures accuracy
PP-OCRv6 is PaddleOCR's latest universal OCR model family, scaling from 1.5M to 34.5M parameters across three tiers, supporting 50 languages. It delivers a +4.6 percentage point improvement in text detection Hmean and +5.1 in recognition accuracy over PP-OCRv5_server. New architecture includes PPLCNetV4 backbone, RepLKFPN for detection, and EncoderWithLightSVTR for recognition. Supports multiple inference backends: Paddle Inference, Transformers, and ONNX Runtime.
Three model tiers: tiny (1.5M), small (7.7M), medium (34.5M) for various deployment scenarios.
Supports 50 languages including Chinese, English, Japanese, and 46 Latin-script languages.
A maintainer of OpenClaw built a system using local open-weight models (Gemma, Qwen) in an agent harness to triage issues and pull requests in real-time, achieving competitive performance with closed models while running on local hardware for minimal cost.
Local models like Gemma and Qwen can effectively classify GitHub issues and PRs for triage.
The system uses an agent harness with a read-only shell (reposhell) to safely inspect code.
Deep research agents that combine private documents with web search can inadvertently leak sensitive information through their query logs. The MosaicLeaks benchmark quantifies this privacy risk and proposes a training method called Privacy-Aware Deep Research (PA-DR) that reduces information leakage by over 3x while maintaining task performance.
MosaicLeaks introduces a benchmark of multi-hop research chains that interleave private local documents and public web queries, measuring three levels of leakage: intent, answer, and full-information.
Standard training for task performance increases both success rate and leakage; training with PA-DR reduces answer/full-information leakage from 34.0% to 9.9% while keeping strict chain success at 58.7%.
LoRA is the most popular parameter-efficient fine-tuning (PEFT) technique, but research shows other methods can outperform it on certain tasks. This article introduces Hugging Face's PEFT library and its benchmarks, discussing how to choose the right PEFT technique based on specific needs, and points out that LoRA is not always the best choice.
LoRA dominates PEFT techniques but may not be optimal.
Hugging Face's PEFT library provides a unified API and benchmarks to help users choose.
A new benchmark harness evaluates the entire process of AI agents using software libraries, using Hugging Face Transformers as a case study. By measuring token usage, time, and error rates across different models and tooling tiers, the authors uncover tradeoffs between ease of use and resource consumption, providing insights for library maintainers and agent users.
Standard benchmarks only check final answers; this harness measures the entire process including token cost and errors.
Three tiers tested: bare install, cloned source, and packaged Skill – each with different overhead.
MolmoMotion is a new 3D motion forecasting model that predicts future 3D point trajectories of objects given a video frame, 3D points on an object, and a language instruction. It outperforms existing methods in robotics planning and controllable video generation. The model is accompanied by the MolmoMotion-1M dataset and PointMotionBench benchmark.
MolmoMotion uses language instructions to guide 3D motion forecasting, outperforming existing methods.
It offers autoregressive and flow-matching variants for deterministic and uncertain scenarios.
AWS's open-source SDK Strands Robots integrates LeRobot, enabling developers to train from Hub datasets and deploy policies on simulated or real robots through a single Agent workflow. This post walks through five steps with a runnable example on a laptop.
Strands Robots SDK exposes LeRobot as composable AgentTools, enabling end-to-end control from dataset to robot hardware.
Simulation and hardware share the same DatasetRecorder and LeRobotDataset format for seamless compatibility.
Z.AI introduces GLM-5.2, a flagship model for long-horizon tasks with a solid 1M-token context, advanced coding capabilities with flexible effort levels, and an open-source MIT license. It achieves top-tier performance on long-horizon coding benchmarks, rivaling closed-source models.
GLM-5.2 delivers a stable 1M-token context for long-horizon engineering tasks.
It leads open-source models on benchmarks like FrontierSWE and PostTrainBench, close to Opus 4.8.