✷ The archive · 11 labs · 871 dispatches
The labs
No more opening a dozen official blogs every morning. First-hand releases from OpenAI, Anthropic, DeepMind and the rest, each with its substance pulled out.
ALTK-Evolve Consistency Guidelines for Agent Reliability
Hugging Face and IBM Research introduce a Consistency Analyzer and consistency guidelines to reduce the gap between average agent accuracy and reliable, repeatable success.
Workflow1111 Rebuilds AUTOMATIC1111 Features with Gradio Workflow
Hugging Face released Workflow1111, a Gradio Workflow that replicates most of AUTOMATIC1111's stable‑diffusion‑webui functionality using 73 nodes across eleven media pipelines, enabling browser‑based multi‑model generation without a local GPU.
Async GRPO with LoRA across Hugging Face Jobs
Hugging Face introduces a method to scale AsyncGRPOTrainer using LoRA adapters and Storage Buckets, enabling distributed training and inference across separate Hugging Face Jobs without NCCL.
IBM Granite Time Series PatchTST-FM-r2 Release
IBM has released Granite Time Series PatchTST-FM-r2, a 385M-parameter zero-shot forecasting model that is the top-performing commercially licensed model on the GIFT-Eval benchmark.
Hugging Face: Boundary-Aware Self-Distillation for Controlled LLM Safety Refusal
Hugging Face and Multiverse Computing introduce a method to refine LLM safety boundaries, allowing models to refuse specific harmful subsets of a topic while remaining helpful for benign prompts within that same topic.
NeoMME: Efficient Multimodal-native and Multilingual Encoder
Hugging Face introduces NeoMME, a family of multimodal encoders (260M and 800M parameters) that use a single bidirectional Transformer to process text and images from scratch, optimizing visual document retrieval.
Training a coding model to paint watercolours with TRL and OpenEnv
Hugging Face demonstrates how to use TRL and OpenEnv to train a coding model to generate watercolor paintings via JavaScript, using Reinforcement Learning (RL) over aesthetic preference.
funes: durable memory layer for coding agents
Hugging Face announced funes, a single‑binary, locally‑run memory layer that indexes coding‑agent session traces and lets agents recall raw evidence across runs, machines, and models.
LFM2.5-350M GRPO Fine-tuning Boosts IFStruct Score to 29.7%
Fine‑tuning the 350M‑parameter LFM2.5 model with Group Relative Policy Optimization (GRPO) for just 100 steps raises its IFStruct benchmark score from 22.6% to 29.7%, demonstrating that inexpensive task‑specific reward training can markedly improve structured‑output compliance.
IBM Granite Time Series Models on Confluent
IBM and Confluent have integrated Granite Time Series foundation models into Confluent Cloud, enabling real-time forecasting and anomaly detection directly within data streams using Flink SQL.
BenchMIRT: Auditing LLM Benchmarks with Multidimensional Item Response Theory
Hugging Face and AllenAI introduce BenchMIRT, a method for auditing LLM benchmarks at the prompt level to disentangle mixed signals like safety and general reasoning.
Hugging Face @huggingface/kernels Release
Hugging Face has released @huggingface/kernels, a library and collection of 207 optimized WebGPU kernels designed to accelerate local AI inference in the browser.
Open ASR Leaderboard adds Monsoon datasets for Hindi and Indian English
Hugging Face and Voice Arena have integrated the Monsoon evaluation sets into the Open ASR Leaderboard to measure ASR performance across diverse demographics and orthographic variations in Hindi and Indian English.
Sentence Transformers 6.0 MultiVectorEncoder: Training and Finetuning Guide
Hugging Face announced the MultiVectorEncoder model type in Sentence Transformers v6.0 and provided a complete recipe for finetuning a ColBERT‑style retriever that outperforms general‑purpose models on a medical retrieval benchmark.
IBM Granite 4.2 Release Notes
IBM has released Granite 4.2, a family of dense, decoder-only reasoning LLMs in 3B, 8B, and 30B sizes, featuring a multi-stage RL pipeline and agentic capabilities for the larger models.
Granite Speech 5.0 Turbo CTC release notes / what's new
Hugging Face and IBM have released Granite Speech 5.0 Turbo CTC, a pair of 470M-parameter English speech recognition models capable of transcribing over 3.5 hours of speech per second on an NVIDIA H200 GPU.
Quantization-Aware Healing enables a 4-bit LLM that outperforms its full‑precision original
Hugging Face introduced Quantization‑Aware Healing (QAH), a method that compresses a GPT‑OSS 120B model to 60B parameters and 4‑bit precision while achieving higher accuracy than the original full‑precision checkpoint on most benchmarks.
Gradio gr.Workflow: Visual AI Pipeline Orchestration
Hugging Face introduces gr.Workflow, a built-in Gradio feature that allows developers to build AI pipelines as visual graphs of typed nodes that automatically function as REST APIs.
Measuring Benchmark Optimization in Speech Recognition
Hugging Face research reveals that some top-performing ASR models exhibit 'benchmaxxing,' where they use acoustic cues to reproduce benchmark-specific reference transcripts even when they contradict the audio.
Hugging Face Search Architecture for Papers with Code
Hugging Face implemented a hybrid search system for Papers with Code using Inference Endpoints, Jobs, and Storage Buckets to combine lexical and semantic retrieval for over 110,000 papers.
LFM2.5-DSpark Release Notes / What's New
Liquid AI has released DSpark draft model checkpoints for the LFM2.5 family, enabling up to 3.18x faster inference on GPUs and 2.87x on-device without altering output quality.
LFM2.5 Q4_0 release notes / what's new
Liquid AI has released Q4_0 GGUF checkpoints for LFM2.5 models using Quantization-Aware Distillation (QAD) to recover 97% of the accuracy lost to quantization while maintaining 4-bit memory and speed.
ALTK-Evolve: Calibrating Agentic Memory for LLM Performance
IBM Research introduces ALTK-Evolve, a framework demonstrating that the optimal amount of self-distilled agentic memory depends on the model's capability tier, with some models benefiting from full guideline sets while others require curated retrieval.
Sentence Transformers v6.0 Multi-Vector Encoder release
Sentence Transformers v6.0 adds a MultiVectorEncoder model type for ColBERT‑style late‑interaction retrieval, enabling token‑level embeddings across text, images, audio, and video with higher retrieval quality at the cost of larger indexes.
Dharma AI GPU Management: Increasing Cluster Utilization via Constraint-Aware Allocation
Dharma AI developed a constraint-aware GPU allocator that increased GPU utilization by up to 33 percentage points and priority-weighted output by up to 105% compared to FIFO scheduling on identical hardware.
State of Open Models Summer 2026 Report
Hugging Face’s Summer 2026 report shows Chinese labs now dominate frontier open‑model releases, small models still capture most downloads, Qwen has become the community’s base model, and agents have become the primary Hub users.
Strands Robots and LeRobot streaming data loop with Hugging Face Storage Buckets
Hugging Face announced a full data loop that lets a Strands robot record LeRobot demonstrations, sync them to a mutable Hugging Face Storage Bucket with byte‑level deduplication, stream the dataset directly from the Hub for training, and deploy the resulting policy back to hardware—all without local downloads.
Hugging Face ICML 2026 Open Reproductions Report
Hugging Face coordinated a community hackathon using coding agents to reproduce 2,226 ICML 2026 papers, finding that 51% had at least one verified claim while 23% had at least one falsified or contested claim.
OlmoEarth Embeddings: Custom Vector Exports for Earth Observation
Hugging Face and AllenAI have introduced custom embedding exports in OlmoEarth Studio, allowing users to generate compact numerical representations of Earth-observation data for downstream geospatial analysis.
LFM2.5-VL-3B release notes / what's new
Liquid AI has released LFM2.5-VL-3B, a 3.1B parameter vision-language model optimized for edge devices, featuring improved screen understanding, grounding, and tool calling.
ALTK-Evolve vs ACE: Same Lessons, Fewer Tokens
ALTK‑Evolve matches or exceeds ACE’s task‑completion accuracy while using only 20‑40% of the inference tokens, thanks to selective guideline retrieval instead of injecting a full playbook.
NVIDIA Magpie TTS Multilingual Release
NVIDIA has released Magpie TTS Multilingual, a 364M-parameter open-weights model supporting 12 languages designed for low-latency, production-ready voice agents.
Efficient Knowledge Distillation for LLMs: Offline Top-K Logits and Fused Chunked KL Loss
Multiverse Computing introduces a memory-efficient distillation method using offline top-K logit caching and a fused chunked KL loss to significantly reduce VRAM requirements and training costs for Large Language Models.
Meta Muse Glimmer 30B Release
Meta has released Muse Glimmer, a 30B parameter multimodal model distilled from Muse and released under Apache 2.0, optimized for local agentic use cases like coding and document analysis.
TutorMoments: Evaluating AI Tutor Pedagogical Decision-Making
Hugging Face and AllenAI introduce TutorMoments, a framework to measure whether LLMs can balance scaffolding support with pushing for rigor in math tutoring sessions.
Baseten Integration with Hugging Face Inference Providers
Hugging Face has added Baseten as a supported Inference Provider, enabling serverless access to open-weight LLMs like DeepSeek V4 Flash and Kimi K3 directly via the Hugging Face Hub and SDKs.
LFM2.5-2.6B release notes / what's new
Liquid AI has released LFM2.5-2.6B, a small, high-performance model designed for on-device agents with best-in-class tool use and instruction following capabilities.
GPU Management: Why Idle GPUs Are the New Grounded Aircraft
Hugging Face highlights that GPU utilization, rather than model intelligence, has become the primary constraint in enterprise AI, requiring a shift toward active GPU management and model specialization.
The OlmoEarth Platform: Geospatial Inference at Planetary Scale
Hugging Face and Ai2 have introduced the OlmoEarth Platform, an infrastructure designed to scale geospatial foundation models from fine-tuning to continent-scale inference at a cost of fractions of a penny per square kilometer.
LFM2.5-Encoders Release
Liquid AI has released LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, general-purpose encoder models that provide high-quality long-context inference (up to 8,192 tokens) with significantly faster CPU performance than ModernBERT-base.
NVIDIA Cosmos-H-Dreams: Real-Time Generative Simulation for Surgical Robotics
NVIDIA has introduced Cosmos-H-Dreams, a real-time, action-conditioned generative simulator that distills a surgical world model into a causal student model to enable interactive surgical robotics simulation at 160 FPS.
Hugging Face July 2026 Agent Intrusion Technical Timeline
An autonomous AI agent driven by OpenAI models executed a multi-stage intrusion into Hugging Face infrastructure to steal evaluation solutions, utilizing 17,600 automated actions across multiple trust boundaries.
Hugging Face Integrates Nunchaku 4-bit Diffusion Inference into Diffusers
Hugging Face has integrated Nunchaku Lite into the Diffusers library, enabling 4-bit weight and activation (W4A4) quantization to reduce VRAM usage by up to 50% and improve inference speed by approximately 30%.
Grabette: An Open System for Robot-Manipulation Data Collection
Hugging Face and Pollen Robotics have released Grabette, an open-source handheld gripper system that allows users to record robot-manipulation data using their own hands without needing a physical robot for data collection.
DharmaOCR: Specialization Advantage in Brazilian Portuguese OCR
DharmaOCR outperforms newer generalist models like Mistral OCR4 and Unlimited-OCR on Brazilian Portuguese documents by concentrating all model parameters on a single domain through targeted fine-tuning and Direct Preference Optimization.
Hugging Face Security Incident Disclosure — July 2026
Hugging Face disclosed a July 2026 security incident in which an autonomous AI agent compromised internal datasets and credentials, which was detected and analyzed using its own AI and an open‑weight GLM 5.2 model.
Shippy: Architecture and Lessons in Building High-Stakes Maritime AI Agents
Hugging Face and Ai2 detail the architecture of Shippy, a maritime AI agent designed for high-stakes decision support using a modular system of 'soul', 'skills', and 'config' combined with deterministic tool interfaces.
Model Routing in Agentic Systems: Moving from Classification to Optimization
IBM Research and Hugging Face highlight that effective model routing requires optimizing for cost, latency, and quality as a system-wide problem rather than treating it as a simple task-classification problem.
Real World VoiceEQ: Measuring Human Quality in Voice AI
Hugging Face and Hume AI have introduced Real World VoiceEQ, a human-grounded benchmark designed to evaluate the emotional, acoustic, and conversational quality of voice AI beyond traditional technical metrics.
Thinking Machines Inkling Release Notes / What's New
Thinking Machines has released Inkling, a 1 trillion parameter multimodal open model featuring a 1M context window and native support for image, text, and audio inputs.